From left to right: University of Vienna researchers Rostyslav Serha, Andrii Chumak, David Schmoll and Sebastian Knauer standing in front of a cryostat. (Courtesy: Ian Ehm)
Bosonic quasiparticles such as magnons show much promise for enabling on-chip quantum information technologies that can be scaled down to the nanoscale, but their too-short lifetime has held back such applications until now. An international team of physicists says it has now increased this lifetime by a hundredfold, thanks to the discovery of a new class of magnons known as short-wavelength dipole-exchange magnons.
Magnons are the collective oscillations of coupled magnetic spins in a material. They are particularly promising for next-generation computing because they could allow for information processing based on the wave-like dynamics of these spins rather than the flow of electrons, thereby significantly reducing energy losses. They can also naturally couple to many other fundamental quasiparticles, such as phonons and photons, which makes them ideal as the building blocks for hybrid quantum systems.
The main drawback of magnons, however, is that they die too fast, explains Andrii Chumak of the University of Vienna, who led this latest study. “Indeed, their lifetimes are typically limited to a few hundred nanoseconds, something that puts a hard ceiling on what we can do with them in a quantum architecture.”
A lifetime of up to 18 µs
In this work, which is detailed in Science Advances, the researchers discovered short-wavelength dipole-exchange magnons in highly pure, single-crystal yttrium iron garnet (YIG) spheres held at millikelvin temperatures. These magnons last for up to 18 µs, which is nearly two orders of magnitude longer than that previously observed for these quasiparticles.
As often in science, the discovery was accidental, says Chumak. “We had prepared three YIG spheres with varying degrees of purity for planned quantum experiments with superconducting qubits, and we simply wanted to characterize these spheres and identify the one that hosted the longest magnon lifetime,” he recalls.
“An 18 microsecond lifetime places magnon coherence on a par with that of typical transmon superconducting qubits used in today’s quantum processors,” he tells Physics World. “This reshapes the role magnons can play in hybrid quantum architectures – from lossy intermediaries to robust quantum memories and low-loss links able to mediate interactions between distant qubits on a chip.”
A programmable on-chip “quantum bus”
And that’s not all: long-lived magnons coupled to superconducting circuits could act as a programmable on-chip “quantum bus”, entangling many distant quantum bits (qubits) made from these hybrid structures along a common waveguide, rather than only nearest neighbours, he adds. The results also point to a clear materials-science route to push magnon lifetimes even further, in particular by continuing to reduce rare-earth impurity concentrations in YIG.
There is a way to go before real-world applications see the light of day, however. For one, the researchers still have some physics to understand, but that is what makes the project so intriguing for them, they say.
Chumak says that the most immediate future step in his team’s work will be an echo-type experiment to measure the lifetime of these magnons directly, rather than inferring it from the so-called parametric instability threshold, as was done in this work. “In parallel, we will need to learn how to excite and detect these short-wavelength magnons efficiently using nanoscale transducers, which is essential before they can be integrated with superconducting qubits.”
“The field of quantum magnonics is still very young compared to other quantum platforms,” Chumak notes, “so it will take some time before we reach actual quantum computing applications – but the path is now much more concrete than it was a year ago.”
Celebrity gossip might break the Internet, but not in the way that quantum computers could. “The advent of quantum computers poses a critical threat, as they could break widely deployed encryption schemes,” warns Lily Chen, a cryptography expert from the US National Institute of Standards and Technology (NIST). Systems at risk include banking encryption, digital signatures, secure messaging, secure shell tunnelling, cryptocurrency and more.
Today’s quantum computers are still too small and error-prone to defeat gold-standard encryption. However, new results from Google Quantum AI and start-up Oratomic suggest that could change, with two widely used cryptographic systems – elliptic curve cryptography (ECC) and the Rivest-Shamir-Adleman (RSA) algorithm – potentially coming under threat sooner than many scientists predicted.
Space–time trade-off
At present, anyone who wants to access encrypted information needs a secret digital key. To obtain this key, an attacker must first solve a difficult mathematics problem. For example, breaking the RSA algorithm boils down to factoring a large number into its prime components. Breaking ECC involves finding a secret number that connects two points on an elliptic curve.
Classical computers might take billions of years to solve these problems. But if an attacker had access to a powerful enough quantum computer, they could solve the problems in mere minutes using an algorithm devised by Peter Shor in 1994.
Several years ago, experts estimated that cracking a typical RSA scheme with 2048-bit keys (RSA-2048) would require tens of millions of physical quantum bits (qubits), which are the building blocks of quantum computers. A year ago, this value dropped to a million. By February 2026 it was down to 100,000. The latest results from California-based Oratomic push the floor even lower, to 10,000 physical qubits. The largest neutral-atom qubit array – realized last year in the lab of Oratomic co-founder Manuel Endres – stands at 6100 qubits. This makes the benchmark of 10,000 feel alarmingly close, though Endres’ array hasn’t yet been used for computation.
The team: Employees at Oratomic, a new neutral-atom quantum computing start-up. (Courtesy: Oratomic)
There are, however, trade-offs. Quantum computers that use fewer qubits or more space-efficient hardware generally have longer computation times. Oratomic’s proposed 10,000-qubit platform would require three years to crack ECC with 256-bit keys (ECC-256) and 120 years to crack RSA-2048. The company’s predicted time-efficient alternative could solve ECC-256 in 10 days, but that would require 26,000 qubits. Solving RSA-2048 in 97 days would take 100,000 qubits.
Oratomic’s numbers have not yet been peer-reviewed, and outside experts say they depend on different assumptions about future hardware developments. “The space-efficient [architecture] is mostly based on assuming aspects that have been demonstrated to work individually in state-of-the-art academic labs,” explains Maria Violaris, a quantum physicist at Oxford Quantum Circuits, who was not involved in the research. “Meanwhile, the time-efficient one relies on more speculative assumptions that need future innovation.”
A second perspective
On the same day as the Oratomic team posted its findings on the arXiv preprint server, researchers at Google Quantum AI released a white paper with their own updated resource estimates. They report that a computer with 500,000 physical qubits made from superconducting circuits could solve ECC-256 in 18 minutes – and potentially even less (see box). Google’s current state-of-the-art processor, Willow, has 105 physical qubits. However, the researchers warn against assuming gradual and predictable progress because quantum computing developments are driven by overcoming scaling barriers rather than by steady increases in processor size.
The quantum threat to cryptocurrencies
Elliptic curve cryptography (ECC) underpins the security of most blockchain networks, including Bitcoin and Ethereum. Bitcoin transactions take an average of 10 minutes, so if a quantum computer can crack ECC and determine the secret key during that window, the transaction could be intercepted and funds stolen in real time.
While Google Quantum AI’s results predict that it would take 18 minutes to solve ECC on a 500,000-qubit quantum computer, they argue that the run time could be effectively shortened in some circumstances. To understand how, imagine planning a heist in which you need to open a safe. Although you won’t know the exact combination until you get your hands on the safe, if you know the model number in advance, you can prepare some tools to help you crack it faster.
A quantum computer could do something similar. According to the Google Quantum AI researchers, half the ECC algorithm only depends on the elliptic curve and not on the specific transaction. A quantum computer could precompute this half, wait in a primed state until a Bitcoin transaction begins, then quickly solve the second half in only nine minutes, dropping below the 10-minute threshold.
Quantum computing platforms that use superconducting, silicon, and photonic qubits are well-positioned for real-time attacks because they tend to compute faster than neutral-atom and ion-based computers. However, the latter could still pose a serious risk through “at-rest” attacks. Such attacks involve adversaries collecting archived and publicly available data, then decrypting it later with few time constraints.
Which threat arrives first will depend on how different quantum computing architectures mature and scale, a path still marked with considerable uncertainty. “Ultimately, feasibility is difficult to say as it depends on how challenging it will be to increase scale or to take a novel approach by engineering [new] hardware,” notes Maria Violaris of Oxford Quantum Circuits.
The high number of physical qubits required for quantum computation comes from the need to detect and correct errors. Google Quantum AI’s estimate is based on a well-known error-correction method known as the surface code. In this approach, physical qubits are arranged in a rectangular grid and interact with their nearest neighbours. Quantum information is spread redundantly across this grid, allowing errors on one physical qubit to be found and fixed. The entire grid is considered one logical qubit, and the ratio of logical to physical qubits is called the encoding rate.
In the surface code, reducing error amounts to adding more physical qubits per logical qubit, and typical encoding rates range from a few hundred to a few thousand. In contrast, the Oratomic team based its estimates on a newer method of error correction called quantum Low-Density-Parity-Check (qLDPC), which reduces error more efficiently by making the physical qubits interact over large distances. Hengyun (Harry) Zhou, a physicist at the Massachusetts Institute of Technology in the US who was not involved in the research, explains that this longer-range connectivity can significantly increase the encoding rate. For qLDPC codes, a typical rate is around 1 to 10, but rates can now go as high as 1 to 2.
Because neutral atoms are highly reconfigurable, neutral atom platforms like those used by Oratomic (and other companies, including QuEra Computing, Infleqtion, Pasqal, planqc and Atom Computing) are naturally suited to the required long-range connectivity that qLDPC codes require. However, Zhou argues that it’s “not completely out of the question” that superconducting qubit platforms could use these codes too. “There is some additional cost that the lack of reconfigurability in those platforms currently leads to, but I would say if we’re thinking about a beyond-10-year timescale, it’s quite imaginable that things could also change for other platforms as well,” he says.
Responsible disclosure
Google Quantum AI’s white paper may represent a turning point in another respect. Rather than being open about their circuit designs, its authors hid them behind a “zero-knowledge proof”, which provided enough information to verify claims while hiding details that they say could provide bad actors with an “instruction manual”.
Superconducting quantum computing: Google Quantum AI’s Willow processing chip. (Courtesy: Google Quantum AI)
This is a relatively novel approach within the quantum computing community, which has thus far followed the conventional academic practice of publishing results with full transparency. A Google blog post expresses hope that “our approach to responsible disclosure can spur an important conversation among quantum computing researchers and the broader public”.
Certainly, it has already spurred a conversation among experts. “This is the first time I’ve ever seen a new mathematical result actually announced that way,” Scott Aaronson, a quantum physicist at the University of Texas at Austin, US, wrote on his blog. “I’m not sure how much it will actually help, as once other groups know that a smaller circuit exists, it might be only a short time until they’re able to find it as well.”
Zhou echoes this sentiment. “These are the kind of results that could potentially have a lot of general societal safety implications, so you want to make sure that they’re safeguarded responsibly,” he observes. “That being said, I think it is also possible that other people, now that they know what is possible, might come up with related constructions.”
What comes next?
In the long run, protecting against threats likely means migrating away from RSA and ECC and towards new mathematical problems that are difficult for both classical and quantum computers to solve. Google recently introduced 2029 as an internal deadline for migrating major system to so-called post-quantum cryptography (PQC), and many experts believe the migration ought to begin now.
“Migrating to PQC is a massive undertaking that won’t happen overnight. Starting migration today is a necessary risk management strategy,” urges Chen from NIST. She notes that NIST has been instrumental in guiding this migration, beginning with its 2016 call for cryptography experts to design and evaluate new algorithms for PQC, and culminating in its publication of the three most promising ones in 2024.
The Google Quantum AI researchers also outline recommendations to help cryptocurrency communities and policymakers prepare for the PQC era. And while urgency permeates their white paper, ongoing PQC efforts prompted them to end it on a positive note. “These trailblazing projects demonstrate that transition to post-quantum cryptography is realistic and instil hope that it will have been completed before the first [cryptographically relevant quantum computers] come online,” they write.
Researchers show Gatekeeper can be bypassed by replacing a previously run legitimate macOS app with malware
Attack requires prior user‑level code execution, then swaps in a malicious app that Gatekeeper won’t re‑verify
Apple dismissed the issue, saying locally rebuilt bundles fall outside Gatekeeper’s scope, leaving risk to social engineering
A pair of researchers claims to have found a way around Gatekeeper, a built-in macOS security feature that helps protect users from running malicious or untrusted software. However Apple doesn’t really see it that way and has seemingly decided not to pursue the issue further.
Gatekeeper’s modus operandi is rather simple - when a user downloads an app from outside the App Store, it verifies the product comes from an identified developer and is notarized by Apple. If it can’t verify it - it won’t allow it to run on the machine.
Now, security researchers Talal Haj Barky and Tommy Mysk claim that, as long as a legitimate app was run at least once on a macOS device, it can be replaced with a malicious version, and Gatekeeper won’t even blink its virtual eye.
Locally built
That also means the attack is not that straightforward to pull off. The threat actor needs to have a way to execute user-level code (for example, a malicious app, a compromised software package installed through a package manager, or a prompt injection attack that tricks an AI agent).
Once that is obtained, they can archive a legitimate app, remove the original, then replace it with malware, and Gatekeeper will not try to re-authorize it. That malicious version can then trick the victim into compromising the device even further, since a certain level of trust was already established.
After reporting the issue to Apple, the company apparently just closed it.
"Apple doesn't consider this attack to be 'modifying' the signed executable," Mysk said. "Instead, Apple says that by archiving/restoring the app bundle, the proof-of-concept code overwrites the entire app bundle, making it locally built. Locally built app bundles are not covered by macOS guards. And this is why access to Keychain or TCC protected directories require system authorization prompts. And for users to accept those is a matter of social engineering attacks that Apple considers out of scope."
It's finally happening! Valve's long-awaited Steam Machine has landed at last, promising to bring gaming-PC-level performance to an ultra-compact console-style device ideal for couch gamers who want to dip their toes into PC gaming for the first time. This is Valve's second run at the console-style PC market, and the gaming giant seems a lot more committed this time around.
And... well, I'm not entirely underwhelmed by it, but it's also not overwhelmingly impressive. I guess you could say I'm distinctly whelmed by the Steam Machine; it's a lovingly crafted piece of hardware that arguably achieves what it sets out to do, but falls down in two extremely important areas.
As you might have already guessed, those areas are performance and pricing. At the default 1080p resolution, it's a capable little machine that should be able to give you a good gaming experience with minimal fuss, but anyone expecting beefy 4K performance on par with the existing slate of top-tier home consoles is going to be disappointed.
Of course, this boils down to what sort of games you want to play. If you've got a 4K TV in your living room and want a compact PC specifically so you can enjoy older games, or the sort of indie game experiences that aren't so freely available on consoles – I'm talking the low-budget viral titles that so frequently dominate TikTok and Instagram Reels – then the Steam Machine might actually be an excellent choice.
Another key aspect of Valve's new console-PC hybrid is SteamOS, the recently updated operating system designed to provide a more console-like experience than conventional platforms like Windows and macOS. It's already available for anyone to put on their own PC, and of course powers Valve's own Steam Deck handheld. In a break from tradition for PC reviews, I've added an extra section to this review to specifically discuss SteamOS, but for now all I'll say is that it gets the job done well.
(Image credit: Future)
Moving on to pricing, it's painfully clear that the ongoing RAM crisis has taken a toll on the Steam Machine's value proposition. Starting at $1,049 / £879 / AU$1,609, it's a fair bit more costly than the PS5 Pro, which my esteemed colleague Rob Dwiar already noted was "too expensive" for the average gamer in his review here on TechRadar - and that was before Sony hiked the price of its home consoles earlier this year.
Considering that the Steam Machine was widely expected to represent an accessible path into the world of PC gaming for those who have previously not ventured beyond the confines of the console market, that price tag is sure to upset some gamers who were eagerly awaiting Valve's return to the console market. Hell, you might even be better off building your own Steam Machine killer.
But it's important to remember that the Steam Machine was probably never going to live up to the price-to-performance ratio of modern home consoles like the PS5 and Xbox Series X, even though it is ostensibly using newer, better hardware. This is due to the game development process and how developers implement platform optimization; if you're making a game for PS5, you're just making a game for PS5. You'll have to take the Pro model's specs into account now too, but at the end of the day you only need to optimize your game's performance for a specific set of hardware components with clearly defined limitations.
On the other hand, optimizing a game for PC means optimizing it for the uncountable millions of possible hardware configurations that exist across pre-built and custom gaming PCs, plus every gaming laptop on the market, plus handheld gaming PCs, and now plus the Steam Machine. It's a wildly different ballgame, and while some developers might skimp on good PC optimization, it's borderline impossible to get a game running consistently well on PC versus a specific console.
Does this entirely excuse the lackluster bang for your buck the Steam Machine offers? No, probably not. It's a mitigating factor that I have taken into account for this review – but quite frankly, I think Valve's second attempt at bringing PC gaming to the masses has slightly missed the mark.
Valve Steam Machine review: Price & availability
(Image credit: Future)
Starts at $1,049 / £879 / AU$1,609
Available now in the US, UK, and Australia
Steam Controller can be added as a bundle
As I mentioned already, the Steam Machine starts at a rather steep $1,049 / £879 / AU$1,609, which gets you the standard configuration with no controller and 512GB of high-speed SSD storage. The actual performance spec doesn't change if you upgrade to the 2TB SSD model for $1,349 / £1,149 / AU$2,109, so you're essentially just paying for the extra storage – and even by current component pricing standards, that's an expensive upgrade.
Whether you pick the 512GB or 2TB model – and despite the price, I'd definitely recommend the latter if you play triple-A PC games, since install sizes have become frighteningly bloated as of late – you get the same semi-custom AMD Zen 4 CPU and discrete AMD RDNA 3 GPU, along with 16GB of RAM and a separate 8GB of GDDR6 VRAM.
Both models can also be bundled with Valve's Steam Controller (which also comes with a charging 'puck' and USB cable), which adds $79 / £59 / AU$119 to the base price of either model. This is actually a pretty decent add-on, since the Steam Controller is a genuinely good gamepad and normally costs $99 / £85 / AU$149 direct from Valve if you buy it separately. Unless you already own one of the best PC controllers, I'd strongly recommend getting the bundle.
It's also worth noting that pre-orders for the Steam Machine sold out almost immediately, so if you want to get your hands on one, you'll have to sign up for the waitlist directly through Steam.
Somewhat limited port selection compared to most PCs
For a product that I found a little disappointing overall, I have to say that I have very little to criticize when it comes to the physical design of the Steam Machine.
Simply put, it looks great. The whole unit is a neat black cube of matte metal measuring roughly six inches along each edge and weighing a little under six pounds, with a thin indentation near the base on the front side, inside which nestles a bright LED lightbar. Beneath this, there's a power button with an additional LED (which stays illuminated while the Steam Machine is in sleep mode) along with two USB-A ports and a microSD card slot for easily adding extra storage. The cube sits atop four small rubber feet, which feel quite durable.
The 17-LED lightbar is more than just a simple power indicator, which I quite like. If your Steam Machine is downloading a game, it 'fills' from left to right to indicate the download's progress, and you can customize it with extra colors and functions in SteamOS too.
(Image credit: Future)
It's significantly smaller than any current-gen home console and the vast majority of pre-built gaming PCs, and certainly a lot smaller than any custom desktop PC the average system-builder could assemble - you'd need to be seriously confident with graphics card disassembly and custom cooling to even get close to something this small.
In fact, it's immediately reminiscent of the Nintendo GameCube I treasured in my teenage years, though I promise I haven't let nostalgia cloud my opinion here. The closest you're likely to get to this sort of compact gaming PC now is probably the Asus ROG NUC that I personally reviewed last year - and while that offers far superior performance, it's also so eye-wateringly expensive that it makes the Steam Machine look cheap.
The front panel is removable and can be easily swapped out with magnetic faceplates. If you spring for the 2TB model, you get two of these included: a red felt fabric one and a rather nice wood-effect one that matches my TV stand perfectly... it looks right at home there, actually. Wink wink, nudge nudge, Valve.
The power supply is (somehow) built into the breathtakingly compact chassis, meaning there's no chunky adapter brick here either - just a regular cable with a two-pin power connector. This is included in the box, along with an HDMI cable (though the rear I/O also features a DisplayPort 1.4 output, which may be preferable if you're connecting it to a desktop monitor rather than a TV).
(Image credit: Future)
That rear I/O is where you'll find the rest of the ports: the aforementioned HDMI and DisplayPort, along with an RJ-45 Ethernet port, two USB-As, and one USB-C. I'll confess that as a fan of physical connectivity, I find this port selection a little lacking, but it's realistically going to be plenty for the target audience of the Steam Machine. It might be less than a fully-fledged desktop PC, but it's comparable to any modern home console. Wireless connectivity uses Wi-Fi 6E and Bluetooth 5.3 – I would've liked to see Wi-Fi 7 here, but it's not a dealbreaker.
The internals are very tidy, with a large air intake grille behind the magnetic faceplate behind which sits the mainboard. A single large upright fan draws air through and out of the rear exhaust. This is (surprisingly) an upgradable system, though it's more along the lines of a gaming laptop than the average desktop PC; you can swap out the M.2 SSD and the RAM, but that's it.
(Image credit: Future)
Lastly, I'll briefly touch on the new Steam Controller. We reviewed it separately here, so I won't go into too much detail, but it's certainly good (if overpriced, like the Steam Machine itself). Pairing the pad and the 'puck' that doubles as a wireless transmitter and charging dock is a seamless process, and the controller itself feels good to use in games. The lack of an audio jack, as seen on other mainline console controllers, is a bit of a shame, however.
I have to admit that I really, genuinely love the design of the Steam Machine. My only real criticism here (and the only thing stopping me from giving it a perfect score in this section) is that it only has one single USB-C port, and it's on the rear. This is a minor gripe, though; for the most part, this is a very well-designed piece of hardware.
Design: 4.5 / 5
Valve Steam Machine review: Performance
(Image credit: Future)
Strong gaming performance at 1080p
Cracks start to show at 1440p; falls apart completely at 4K
No resolution upscaling or frame generation modes are used unless otherwise specified.
Benchmarking the Steam Machine was... interesting, to say the least. I had hoped to mostly stick to our usual suite of tests used for benchmarking the best gaming PCs and best gaming laptops, but the vast majority of those systems are running on Windows 10 or 11. SteamOS, while a very solid platform these days, simply doesn't have the same degree of flexibility as Microsoft's flagship operating system.
For starters, the Steam Machine locks itself to a maximum framerate based on the refresh rate of the monitor it's connected to, which meant I had to seek out a different high-refresh-rate display to get accurate results. Notably, the Steam Deck does something similar with its built-in 60/90Hz screen, so I suspect that this is a SteamOS quirk rather than any hardware-imposed limitation, but I wasn't able to find any way to disable it.
Then I got to the compatibility issues. Half the tests we usually run (most pertinently Geekbench and 3DMark) simply can't be installed and run natively in SteamOS's default Big Picture Mode, so I had to cut down my testing selection fairly significantly. To compensate (and to provide a better idea of how well the Steam Machine can actually handle gaming), I expanded my gaming tests to include more graphical configurations than usual.
Furthermore, you have to dig into the settings of SteamOS itself to uncap the resolution, which is set to FHD 1080p by default. Eagle-eyed gamers clocked prior to the Steam Machine's release that Valve had quietly lowered expectations by adjusting the language used in the product's store page from "4K gaming at 60 FPS" to "Up to 4K gaming with FSR", crushing hopes that it would be a true competitor to the likes of the PS5 Pro. Clearly, 1080p is the ideal target, which will no doubt be a disappointment for some potential customers.
(Image credit: Future)
Anyway, let's talk about the actual performance. At 1080p with graphical settings maxed out, the Steam Machine can handle almost anything you throw at it; the discrete semi-custom AMD RDNA 3 GPU inside this little black box was able to crack 60fps in most games I tested at FHD resolution (sometimes requiring the use of AMD's FSR upscaling tech, though this had virtually zero impact on visual fidelity), with the exception of titles that use ray tracing.
With ray tracing turned off, you'll be fine; turn it on, and you can expect framerates to tank even if you also switch on FSR. Cyberpunk 2077 looked and played great at 1080p Ultra settings with no upscaling, but switching to the Ray Tracing Ultra preset (with AMD FSR 3 set to Balanced) saw my average framerate drop to a little over 30 in the built-in benchmark, while actual in-game performance was even choppier.
Annoyingly, some games don't seem to properly support the most recent version of FSR on SteamOS; while I've been able to use it in games like Civilization VII and Shadow of the Tomb Raider while testing Windows-based gaming laptops and PCs, FSR 3 was conspicuously absent from the settings menu of both games on the Steam Machine - Civilization would only let me select FSR 1, while Shadow had no upscaling options available at all.
(Image credit: Future)
Once I started to crank things up to 1440p and 4K resolution, though, the cracks started to show. Older titles (and 2D games, like my current obsession, Slay the Spire 2) largely run fine at higher resolutions - but unfortunately, I could probably say the same about any home console and plenty of cheaper desktop PCs.
Modern triple-A games start to struggle a little at 1440p, but remain playable provided you're willing to tone down the settings a bit. I've only included the full suite of 1080p and 4K results in the results box above, but to give you a few quick examples: Cyberpunk 2077 remained playable with an average framerate of 68fps at 1440p Ultra with FSR 3 set to balanced (no ray tracing), while Total War: Warhammer III dropped to 45fps at Ultra - but could be salvaged to a stable 63fps with the Medium graphical preset.
At 4K, don't expect to be playing the latest hit games at high framerates; at least, not without some very significant graphical tweaks. The now eight-year-old Shadow of the Tomb Raider was the only game I tested that made it above 60fps at 4K resolution, and that was with the graphical settings turned down significantly.
As for more demanding titles, just don't bother. Total War: Warhammer III barely made it above a 30fps average at Low settings, with frequent drops into sub-30 territory, while Cyberpunk 2077 gave abysmal benchmarking at anything above the lowest possible graphical settings, even with FSR upscaling turned on, and struggled heavily to maintain a consistent framerate during actual gameplay. Turning on FSR frame generation did technically give a stable 30fps at 4K Ultra, but with the added issue of some visual tearing and artifacting, especially during busy scenes.
(Image credit: Future)
I also experienced a small number of crashes during my testing process, specifically in Total War: Warhammer III and Marathon. It's possible that these were merely down to game instability rather than any specific fault with the Steam Machine or SteamOS (I've personally had Marathon crash at least once on my Windows 10 gaming desktop), so I won't mark it down for this, but I've opted to include it for the sake of comprehensiveness.
Lastly, I'd like to briefly discuss the thermal performance of the Steam Machine, which is a shining beacon amidst the otherwise fairly underwhelming overall performance. The large cooling fan at the rear of this little cube was whisper-quiet throughout all my testing, and the casing barely got more than a little warm. Whatever thermal wizardry Valve pulled off here, it clearly works.
Performance: 3 / 5
Valve Steam Machine review: SteamOS
SteamOS 3 is a very solid operating system
Accessible for newcomers, but still packs advanced features
Optional Linux desktop mode
A bonus section appears! Yes, I'm going to take a moment here to discuss SteamOS, since it's arguably the linchpin of the Steam Machine's whole 'console rival' mission statement. Does it really make PC gaming more accessible? I'm pleased to say yes, it certainly does.
An Arch Linux-based operating system, SteamOS is now in its third main iteration, and it's improved greatly from its humble beginnings as the full-screen version of the Steam desktop app. It's easy to set up and offers a pleasingly straightforward interface that even a gaming rookie should be able to navigate, while also packing the full suite of settings found in the Windows desktop version of Steam.
One nifty feature is the optional Linux desktop mode. It's quite effective, if basic, letting you install Linux-compatible programs like LibreOffice and the Firefox browser. It's neat, but it's a feature that will likely only be used by a small subset of buyers since it doesn't really offer anything for pure gaming users.
(Image credit: Future)
Naturally, you can connect a keyboard and mouse as well as a controller, which is useful for games where typing is necessary, but it turned out to be even more vital. I wasn't actually able to complete the SteamOS setup without first connecting a mouse to select the system language, and when playing some other games after my initial benchmarking process, I found that I couldn't complete the initial setup for Forza Horizon 5 using just the Steam Controller, which was annoying.
As effective as SteamOS is, compatibility issues are an inevitability when dealing with a Linux-based system. While the vast majority of games on Steam are fully compatible, there will be specific outliers, and there are plenty of non-gaming programs that aren't available. And of course, you can't play games from other storefronts like Epic and GOG, unless you decide to go the whole hog and install Windows 11 on it instead - which is an option!
Should I buy the Valve Steam Machine?
Valve Steam Machine: Scorecard
Attributes
Notes
Rating
Value
The Steam Machine is an expensive product for the specs and experience it offers, to the point where it struggles to compete with mainstream home consoles.
2.5 / 5
Design
A durable, high-quality chassis with an impressively compact design. The port selection is decent for a console-style device, and the swappable faceplates add a bit of visual flair.
4.5 / 5
Performance
1080p performance is great, but triple-A games start to struggle at 1440p and 4K. Considering the price, the gaming quality is decidedly lackluster, and you may encounter some compatibility issues with SteamOS.
3 / 5
Total Score
I really want to love the Steam Machine, and it arguably achieves its goal of competing with living room consoles, but the asking price is just a bit too high for what it offers.
3.33 / 5
Buy the Valve Steam Machine if…
You want living-room PC gaming The compact design and straightforward functionality of the Steam Machine make it great for a couch gaming setup.
You're looking to make the jump to PC gaming Want to ditch your PS5 in the wake of Sony's new digital-only game policy? The Steam Machine offers a very accessible way to enter the PC gaming space without the complexities of PC-building.
Don't buy it if…
You're on a budget While you could easily spend a lot more on a compact gaming PC, the price of the Steam Machine is unfortunately a little too high to be appealing.
You want to play games from multiple storefronts The best experience you'll get here is by sticking with SteamOS, but that will lock you out of playing games from other storefronts like Epic and Prime Gaming.View Deal
Valve Steam Machine review: Also consider
Alienware Aurora It's a bit more expensive than the Steam Machine, but if you want a pre-built gaming PC that can comfortably tackle 1440p and potentially even 4K gaming, the more affordable configurations of the Alienware Aurora are a solid choice. Read our full Alienware Aurora review.
Asus ROG NUC If money is no object, the latest Asus ROG NUC gaming PC is unquestionably the one to pick if you want a compact system for a living-room gaming setup. This NUC probably has the best performance-to-size ratio of any gaming PC on the market right now. Read our full Asus ROG NUC review.
How I tested the Valve Steam Machine
(Image credit: Future)
Ran multiple PC game benchmarks
Connected to my TV to use for 'couch gaming'
8+ years of experience reviewing dozens of laptops and PCs
My review process for the Steam Machine was a little different from how I'd usually approach a gaming PC or laptop review, due largely to SteamOS's Big Picture Mode not supporting our non-gaming tests.
Still, I ran the Machine through all the usual gaming tests, then took it downstairs to my living room and set it up on my TV stand to use it for some good old-fashioned couch gaming. As someone who was raised on Nintendo consoles but made the jump to PC gaming in my teens, I was keen to get the full intended experience out of the Steam Machine.
In addition to our usual game tests, I played a bunch of other games over a full weekend spent rigorously testing (read: having lots of fun with) the Steam Machine. These included Marathon, Warframe, Ratchet & Clank: Rift Apart, and the frighteningly addictive Slay the Spire 2.
I've been testing and reviewing computer hardware for almost a decade, and I've been a gamer for much longer than that. Before joining the TechRadar team, I worked at the PC-building hobbyist magazine Maximum PC, so you can trust that I'm putting every ounce of knowledge and experience at my disposal into bringing you high-quality, insightful, and non-biased reviews.
Recorded Future found an Iran-linked group spreading spyware
The malware is delivered through fake VPN and media player apps
Researchers assess that most targets are Iranian users
A new report from Recorded Future's Insikt Group describes a campaign that inverts the whole point of a privacy tool: fake VPN apps built specifically to spy on the people who install them.
Researchers have linked fresh infrastructure to an Iran-nexus threat cluster they track as TAG-182, which is using fake VPN and media player downloads to allegedly deliver a surveillance tool called MarkiRAT. The group is "highly likely" to be targeting Iranians living inside and outside the country, the report says.
It's a blunt reminder that choosing one of the best VPN services is a lot more secure than downloading free, unvetted tools.
Fake apps, real surveillance
Insikt Group identified a cluster of attacker-controlled domains allegedly used to stage downloads of applications that appear nowhere on Google Play or Apple's App Store.
Two names stand out: Pis2ray VPN and a media player branded YESHICA, which was quietly renamed YESHICA YEPlayer in March 2026 after researchers publicly exposed the original.
According to researchers, if you download and run one of these files, you get MarkiRAT, a remote access Trojan. In plain terms, that is software that hands control of your device to somebody else.
A fake VPN app. A fake media player. Both delivering Iranian government surveillance #malware to targeted dissidents. Insikt Group has new research on TAG-182 and MarkiRAT:https://t.co/G7p9JO6peT#ThreatIntelligence #Cybersecurity pic.twitter.com/GwDyvGC99rJuly 2, 2026
Analysts have documented it capturing screenshots and uploading them to attacker-run servers, while disguising itself under believable process names.
It also abuses BITS, the background service Windows uses to fetch updates, to pull down further files. Because that activity looks like ordinary system housekeeping rather than an attack, it tends to slip past routine cleanup.
MarkiRAT is not new. It has previously been used by Ferocious Kitten, a group Kaspersky documented conducting years of covert surveillance against activists inside Iran.
Recorded Future stops short of attributing TAG-182 to any specific Iranian agency, but places it within a broader ecosystem of state-aligned surveillance groups.
Why a fake VPN makes such an effective lure
Distribution runs largely through social media. Insikt Group found Instagram posts promoting Pis2ray VPN in the weeks following street protests in Iran in late 2025, and again around the country's prolonged internet shutdown, which ended with partial restoration of access on 26 May 2026.
The people most desperate for a virtual private network (VPN) in a censored country are exactly the people most likely to install one from a social media link, because the official stores are often the very thing they cannot reach.
Recorded Future considers it almost certain that most targets are located in Iran or tied to anti-government movements in Europe and North America. TechRadar has covered previous Iran-linked fake VPN campaigns, and this one seems to follow the same pattern with better infrastructure.
How to stay safe
Most readers will never be targeted by a state actor, but the underlying lesson travels.
Install VPN apps only from official stores, and check that the provider has a real, verifiable presence outside the app listing.
Treat any VPN promoted through an Instagram post, a Telegram channel, or a direct message as suspect, however polished it looks.
Star ratings are a weak signal, since fake reviews are cheap.
Quantum materials could transform technologies ranging from powerful computers and highly secure communications to advanced energy systems, but one major obstacle has long stood in the way. Most quantum materials reveal their unusual properties only when chilled to temperatures near absolute zero. At room temperature, continual atomic vibrations caused by heat can overpower the fragile [...]
A new NYT Strands puzzle appears at midnight each day for your time zone – which means that some people are always playing 'today's game' while others are playing 'yesterday's'. If you're looking for Friday's puzzle instead then click here: NYT Strands hints and answers for Friday, July 24 (game #873).
Strands is the NYT's latest word game after the likes of Wordle, Spelling Bee and Connections – and it's great fun. It can be difficult, though, so read on for my Strands hints.
Want more word-based fun? Then check out my NYT Connections today and Quordle today pages for hints and answers for those games, and Marc's Wordle today page for the original viral word game.
SPOILER WARNING: Information about NYT Strands today is below, so don't read on if you don't want to know the answers.
NYT Strands today (game #874) - hint #1 - today's theme
What is the theme of today's NYT Strands?
• Today's NYT Strands theme is… Frocks and smocks
NYT Strands today (game #874) - hint #2 - clue words
Play any of these words to unlock the in-game hints system.
WORM
RINK
RIDE
PAIN
MUTE
LUMP
NYT Strands today (game #874) - hint #3 - spangram letters
How many letters are in today's spangram?
• Spangram has 13 letters
NYT Strands today (game #874) - hint #4 - spangram position
What are two sides of the board that today's spangram touches?
• First side: top, 3rd column
• Last side: bottom, 5th column
Right, the answers are below, so DO NOT SCROLL ANY FURTHER IF YOU DON'T WANT TO SEE THEM.
NYT Strands today (game #874) - the answers
(Image credit: New York Times)
The answers to today's Strands, game #874, are…
GOWN
TUNIC
PINAFORE
KIMONO
DIRNDL
MUUMUU
SPANGRAM: YOUWEARITWELL
My rating: Hard
My score: 1 hint
THis was a game that started off easy and ended baffling.
After collecting the obvious pieces of clothing in GOWN, TUNIC, and PINAFORE, what remained on the board was a letter soup with no obvious words at all. After piecing together the spangram I did see KIMONO, but then drew a blank.
I needed a hint to get DIRNDL — the name for an Alpine dress — which left me MUUMUU, another garment I had not previously heard of but now know, thanks to Wikipedia, is a loose dress of Hawaiian origin.
Yesterday's NYT Strands answers (Friday, July 24, game #873)
EXTRA
SURPLUS
EXCESS
LEFTOVER
SPARE
UNUSED
SPANGRAM: ANDTHENSOME
What is NYT Strands?
Strands is the NYT's not-so-new-any-more word game, following Wordle and Connections. It's now a fully fledged member of the NYT's games stable that has been running for a year and which can be played on the NYT Games site on desktop or mobile.
I've got a full guide to how to play NYT Strands, complete with tips for solving it, so check that out if you're struggling to beat it each day.
A new Quordle puzzle appears at midnight each day for your time zone – which means that some people are always playing 'today's game' while others are playing 'yesterday's'. If you're looking for Friday's puzzle instead then click here: Quordle hints and answers for Friday, July 24 (game #1642).
Quordle was one of the original Wordle alternatives and is still going strong now more than 1,500 games later. It offers a genuine challenge, though, so read on if you need some Quordle hints today — or scroll down further for the answers.
Enjoy playing word games? You can also check out my NYT Connections today and NYT Strands today pages for hints and answers for those puzzles, while Marc's Wordle today column covers the original viral word game.
SPOILER WARNING: Information about Quordle today is below, so don't read on if you don't want to know the answers.
Quordle today (game #1643) — hint #1 — Vowels
How many different vowels are in Quordle today?
• The number of different vowels in Quordle today is 4*.
* Note that by vowel we mean the five standard vowels (A, E, I, O, U), not Y (which is sometimes counted as a vowel too).
Quordle today (game #1643) — hint #2 — repeated letters
Do any of today's Quordle answers contain repeated letters?
• The number of Quordle answers containing a repeated letter today is 1.
Quordle today (game #1643) — hint #3 — uncommon letters
Do the letters Q, Z, X or J appear in Quordle today?
• No. None of Q, Z, X or J appear among today's Quordle answers.
What letters do today's Quordle answers start with?
• S
• S
• B
• G
Right, the answers are below, so DO NOT SCROLL ANY FURTHER IF YOU DON'T WANT TO SEE THEM.
Quordle today (game #1643) — the answers
(Image credit: Merriam-Webster)
The answers to today's Quordle, game #1643, are…
SALON
SHEEN
BURST
GOURD
A nice and easy game (for me at least) to start the weekend.
My starter words paid huge dividends, giving me two straightforward anagrams to solve and making significant headstarts with SALON and GOURD. Happy days.
Daily Sequence today (game #1643) — the answers
(Image credit: Merriam-Webster)
The answers to today's Quordle Daily Sequence, game #1643, are…
A new option for the Copilot key has been spotted in testing
Windows 11 users are getting the ability to remap it to 'Do nothing'
This comes on top of work already underway to redefine the key to Right Ctrl or the context menu
Microsoft's latest revolutionary idea for Windows 11 is to have a key on your keyboard which can be remapped to do absolutely nothing.
Windows Latest noticed the latest twist in the tale of the Copilot key, as highlighted by regular leaker PhantomOfEarth on X, who spotted the development hidden in a recent Windows 11 preview build.
In the choices to customize the Copilot key – which, by default, summons Windows 11's Copilot AI (of course) – there's a new option to simply 'Do nothing'.
Yes, that's correct – you can change the key to simply be a pointless piece of plastic that does precisely zip. (Or at least you'll be able to in testing, when this ability comes into play, that is, although some Windows 11 testers have reportedly seen it live in preview already).
As you may have seen earlier this year, Microsoft is working on bringing fresh options to the Copilot key that include customizing it to invoke the context (right-click) menu, and to turn it back into Right Ctrl (which was the key that the Copilot button replaced on Copilot+ laptops and some keyboards).
Analysis: do nothing, be happy
(Image credit: rui vale sousa / Shutterstock)
What's the point of a key that does nothing? Good question. Maybe it's the keyboard equivalent of a stress ball – you can hammer on that key as much as you want when you're really annoyed, and it won't splurge a load of characters into the document you're working on as you exorcise your frustrated demons.
Okay, so maybe not, but perhaps the idea is that this is aimed at people who really hate the Copilot key, but haven't decided what they might want to remap it to. So, they can go ahead and elect for the key to do nothing if it's accidentally hit, while they figure out what they might want to redefine in its place. That's stretching things, admittedly, but it's the best guess I have.
At any rate, it's a harmless enough option, although it feels inescapably odd to have a 'blank' key on your keyboard. How far we've fallen since the Copilot key was first revealed and trumpeted by Microsoft as the biggest thing to happen to keyboards in 30 years since the Windows key was introduced. And indeed more recently, Microsoft promoted the Copilot key as the "button you can press to fix everything", so there's some irony that it could soon (optionally) become the button you can press to fix 'nothing' in a very literal sense.
Don't forget that you don't have to wait for Microsoft's Copilot key remapping functions to arrive in the release version of Windows 11. You can use the Keyboard Manager in PowerToys to do this instead, although it'll be far neater to have these options in Windows 11's settings, of course.
Razer is a household name in the webcam space, with its Kiyo brand being one of the most popular choices for streamers and PC gamers. The latest Kiyo V2 is another sterling example of Razer's commitment to premium build quality and high-end components, and for the most part, it's a great product.
Delivering either 4K recording at 30 frames per second or 1080p at 60, the Kiyo V2 captures quality video with a high degree of customization via the Razer Synapse software. It's fairly pricey (don't expect to see it on our list of the best cheap webcams), but for a 4K webcam with this sort of build quality and feature set, it's actually quite competitively priced.
(Image credit: Future)
Coming in black, white, and pink colorways (mine is the pink 'Quartz' model), the Razer Kiyo V2 is an extremely well-constructed product. It's durable and pleasingly tactile. The base can be extended and rotated to capture the perfect angle, and it easily clasps onto any monitor; I tested it both with the two monitors in my home desktop setup and my laptop, and it snapped into place with no issues. It's a little heavy, however, so laptops with flimsier display hinges might struggle.
The design is a departure from previous Razer webcams, which typically opted for a singular large casing surrounding the lens construction. Instead, it's a smaller lens housing with an oblong body, immediately reminiscent of the Logitech MX Brio 705.
Not everything has been abandoned with this new design, however. The lens housing, like previous Kiyo models, includes a pleasingly tactile rim that can be twisted to close the built-in physical privacy shutter. I'm always a fan of privacy solutions like this, though given the large size of the glass lens face, it would've been nice to see a lens cap included to keep it dust- and scratch-free.
(Image credit: Future)
It would potentially be nice if the webcam could be detached from the attached mount in order to fit it to a different mount; many content creators will have their own video setups and may want to use their own camera stands. However, the mount does have a threaded screw connector on the base, letting you fit it atop most industry-standard camera mounts.
In terms of video quality, it's great. 4K video is sharp and colorful, and even zoomed in, I found I looked perfectly detailed in the footage I recorded. Color reproduction is great, and the camera performed well in a range of light conditions. Low-light performance resulted in some mild oversaturation, but I've seen worse performance in more expensive cameras. It also supports HDR, though only at 30fps even at lower resolutions.
The autofocus works well, but I found it was a little slow to adjust if I quickly moved closer or further from the camera. Having tested quite a few webcams in my time, I feel that a webcam in this price range should be a bit snappier to focus.
That big lens means that the default field of view is huge. A 93-degree FOV means that the Kiyo V2 is liable to capture not just yourself but the entire room you're in, but thankfully it's easy enough to tweak this via the zoom and crop tools in Razer's Synapse software.
(Image credit: Future)
On the topic of Synapse, it's still a frankly rather clunky piece of software, but for webcam users who want granular control over their hardware, it does a solid job. In addition to zoom and crop tools, you can also manually tweak the autofocus, auto exposure, and color balance.
The Kiyo V2's software-side feature set also includes a selection of AI-powered tools, which are covered prominently in the marketing material for the webcam. These include dynamic light correction and noise reduction, but these features already existed in Synapse, and to be honest, I'm not entirely clear on what AI brings to the table here.
(Image credit: Future)
Look, Razer: I get it. I'm sure there's tremendous pressure from your shareholders to make everything AI-powered, because it's the buzzword du jour. But anyone who actually wants AI in their webcam (for whatever inane reason) likely isn't going to be better served by the Kiyo V2 than any other cutting-edge 'AI' webcam, like the Obsbot Tiny SE.
One element that does actually benefit from the involvement of machine learning is the auto-framing feature, which uses AI-powered head-tracking to keep you in the center of the frame even if you move around. However, since the Kiyo V2 lacks a mechanized gimbal, it can't alter the frame while fully zoomed out and can only track you so far; it can keep you in center frame if you move by a foot or two, but it won't follow you around a room.
On the bright side, you don't actually have to use Synapse and the AI features at all. The Kiyo V2 works great straight out of the box; just plug in the bundled cable, and you're off to the races, assuming you're happy with the default wide-angle FOV.
(Image credit: Future)
Now, for the only real problem I have with the Razer Kiyo V2. It comes with a USB-C to USB-C cable in the box (the one with my review unit is pink, like the webcam itself), and as far as I can tell, you have to use this cable to access the full 4K functionality of the camera.
Considering that the cable is USB-C-exclusive, you'll need a free USB-C 3.0 port to make full use of the Kiyo V2. I tested it out with a USB-C to USB-A adapter, as well as using one of my own USB-C cables; no dice. It's frustrating, because USB-C ports can be in short supply for users connecting multiple peripherals, and the bundled cable is a bit short and inflexible.
(Image credit: Future)
Lastly, the built-in microphone is... fine, I guess. It's good enough for video calls, but anyone aiming to produce professional-level video content should definitely use a proper standalone microphone.
It's fair to say that the Razer Kiyo V2 deserves a spot among the best webcams on the market. There are a few niggling issues, but nothing that should ruin the experience for a gamer who wants a high-end webcam. For hardcore content creators, it remains a sound choice, but you should be aware of the downsides.
Razer Kiyo V2 webcam: Price & availability
How much does it cost? $149.99 / £149.99 / AU$279.95
When is it available? Available now
Where can you get it? The US, UK, and Australia
At $149.99 / £149.99 / AU$279.95, the Razer Kiyo V2 is significantly cheaper than the brand's own older Kiyo Pro webcam. If you're looking for a high-quality 4K cam, it's not an unreasonable price.
Sure, you could easily get a cheaper 4K webcam, but you'll need to risk it on a less established brand, and you almost certainly won't be getting the same rich feature set seen in the Kiyo V2. Similarly effective models from other premium brands like Logitech and Elgato will cost a comparable amount, so I'm pleased to say that despite its reputation for pricey hardware, Razer got the price right this time around.
Razer Kiyo V2 webcam: Specs
Razer Kiyo V2
Price:
$149.99 / £149.99 / AU$279.95
Supported resolutions:
4Kp at 30fps, 1440p at 60fps, 1080p at 60fps
Sensor:
8.3 MP Sony STARVIS
Field of view:
93 degrees
Connection:
USB-C
Features:
Digital zoom, AI auto-framing, AI autofocus, dynamic color correction, HDR, mechanical privacy shutter
Software Integration:
Razer Synapse
Should you buy the Razer Kiyo V2 webcam?
(Image credit: Future)
Razer Kiyo V2 webcam report card
Value
The Razer Kiyo V2 is still fairly expensive, but it's reasonably priced for a high-end 4K webcam and much better value than the previous Kiyo Pro.
4 / 5
Design
A fresh chassis design that retains the excellent physical privacy shutter of older Kiyos, but the forced USB-C is annoying.
4 / 5
Performance
Audio is merely decent, but the picture quality and color balance are fantastic and Razer Synapse packs a ton of extra features.
4 / 5
Avarage rating
A high-quality webcam with a broad feature set and more competitive pricing, the Razer Kiyo V2 is a great product with only a few small flaws.
4 / 5
Buy it if...
You want 4K At this price point, the Kiyo V2 is competitive within the 4K webcam space - you could find cheaper ones, but only if you're willing to accept a drop in quality.
You need wide-angled recording If you're aiming to capture a larger space (or multiple video participants), the 93-degree FOV of the Kiyo V2 is perfect.
Don't buy it if...
You're short on ports Since it needs to use the bundled USB-C cable to access full 4K recording, the Kiyo V2 is a poor choice for anyone who already has lots of ports in use on their PC.
You're on a budget At $149.99 / £149.99 / AU$279.95, the Kiyo V2 is still pretty darn expensive, even if it's far better value for money than the older Kiyo Pro.
Razer Kiyo V2 webcam: Also consider
Insta360 Link 2C Our current top pick of the 4K webcams available on the market right now, the Link 2C offers superior low-light performance and autofocus while also offering the same high-quality video and AI features as the Kiyo V2.
Obsbot Tiny 2 The Tiny 2 from Obsbot delivers 4K video along with AI-powered tracking via its mechanical gimbal mount. For those who want a high-end video experience, this is a great pick.
8+ years of experience reviewing computer hardware
I replaced my usual webcam (the Insta360 Link 2 4K) with the Razer Kiyo V2 for two weeks, using it for all my normal video calls during the workday. I also tested it in various different lighting conditions by recording video separately and comparing the clips side by side. Naturally, I installed (well, reinstalled) Razer Synapse to properly test out all the software-side features as well.
I've been testing and reviewing PC hardware for more than eight years, starting at the PC-building hobbyist magazine Maximum PC before making the leap to TechRadar's computing team. I put all of my knowledge and experience behind each and every one of my product reviews, so you can be assured that you're getting an informed and bias-free judgement.
Model Context Protocol (MCP), the emerging standard for connecting AI models to external tools and enterprise data, is undergoing its biggest architectural overhaul yet.
The latest release candidate, scheduled for release on July 28, removes protocol-level sessions in favor of a stateless architecture, a change which industry experts say is intended to make MCP easier to deploy across standard cloud infrastructure as enterprises move AI pilots into production.
“The session-based model made sense when MCP servers were local processes on a developer’s laptop. In production, it became an operational tax,” said Muskan Bandta, cloud associate at ZopDev.
“When your infrastructure team asks whether MCP services can scale like other cloud applications, the answer used to be ‘not quite.’ With the move to a stateless architecture, the answer is now yes,” Bandta added.
Earlier versions of the protocol maintained information about every client connection, meaning servers had to keep track of each session throughout an interaction. While that approach worked well for local development, it complicated deployments across multiple servers because requests often had to be routed back to the same machine, limiting scalability and making MCP a less natural fit for modern cloud architectures.
“Under the new stateless design, every request contains the information needed for any available server to process it independently. Applications that need to maintain context across multiple requests can still do so, but developers must now manage that state explicitly rather than relying on the protocol itself,” she said.
This transition to a stateless design goes beyond simplifying infrastructure by fundamentally changing how AI applications manage and share context across tools, according to Amit Jena, AI development manager at IT consulting firm Kanerika.
Instead of keeping application state hidden inside protocol sessions, the new design makes it explicit, allowing AI models to access, reason over, and pass that information between tools, giving developers greater control over how context is preserved and shared across tools, Jena said.
It should also make AI workflows more portable, resilient, and easier to orchestrate across distributed environments, he said.
MCP’s new features
Other changes to MCP include the addition of a Multi Round-Trip Requests (MRTR) mechanism that changes how AI agents request additional information they need to complete a task.
Instead of relying on a persistent connection between the client and server throughout the interaction, the new mechanism lets the server request additional input through a standard request-response exchange before continuing the task, Jena said.
Routable transport headers, another addition, enable API gateways and other networking infrastructure to identify and route MCP requests without inspecting their contents.
They reduce processing overhead, lower latency, and let enterprise teams enforce routing, rate-limiting and security policies more efficiently using existing API management infrastructure, Jena said.
MCP is also getting an updated authorization framework built around OAuth 2.1 and OpenID Connect; interactive MCP Apps; and deterministic caching of tool and resource listings to improve LLM prompt-cache hit rates, potentially saving on token costs.
Rebuilding the trust boundary
The MCP release steering committee also decided to deprecate some legacy features, including Roots, Sampling, Logging, the older HTTP+SSE transport and Dynamic Client Registration, although these will continue to work in this version and any other released over the next year.
The deprecation of Sampling is likely to have the biggest impact because it changes who is responsible for interacting with foundation models, said Jena.
“Sampling let MCP servers invoke the LLM through the client, which meant the server had a callback path into the model without owning that connection. Deprecating it means rebuilding that trust boundary,” Jena said. “Your server now calls the model provider directly. That changes your network architecture, your auth model, and depending on how you’ve built cost attribution, your billing flow.”
The year-long transition period will be enough for teams to audit their sampling dependencies now, said Jena: “The risk is that teams who haven’t implemented sampling themselves won’t know if a third-party MCP server they’re depending on uses it.”
Updated MCP SDKs
To accompany the protocol update, there are updated MCP SDKs for Python, Typescript, Go, and C#. These support both the old and new protocol versions, so new clients can continue communicating with older servers, while updated servers will also support older clients, reducing the risk of immediate disruptions.
That backward compatibility should make the transition largely incremental, except for enterprises that built custom infrastructure around MCP’s earlier session-based architecture, Bandta said.
Identifying and auditing those session dependencies may not be easy, Jena warned.
“Session management complexity tends to be hidden across multiple layers — the gateway config, the deployment scripts, the monitoring dashboards. The code change is small; finding everywhere the assumption lives is what takes time,” he said.
Researchers have discovered asecurity flaw in AI web browsers that could result in users having their data exposed by malicious websites.
Browsers equipped with AI agents, such as ChatGPT's Atlas, are like ordinary internet browsers but with additional chatbot functionality baked into the software. Using AI, agentic browsers can summarize websites, search for specific information, and even automate repetitive tasks.
For example, somebody may use an AI browser to make a purchase — whereas they would have to browse a webstore for the item and then make the payment if they used a conventional browser.
Many agentic browsers have only been released in 2025, and they're already rising in popularity. Scientists, however, have warned in a new study that many popular AI browsers bypass an important security measure that keeps their data private. They presented their findings April 26 at the Agents in the Wild Workshop in Rio de Janeiro, Brazil.
"Browser agents aren't ready for the public," said co-author of the study David Kohlbrenner, an assistant professor of computer science and engineering at the University of Washington, in a statement.
"Even if you’re a relatively savvy user, if these agents have access to a browser that contains your credentials — your email, your bank account, whatever it is — you should not trust that these systems are ready to truly protect your information. They may get there in time, but they're not there yet."
Bypassing embedded security
Conventional internet browsers use a security protocol called the "same-origin policy," which ensures that multiple websites a user is visiting at the same time do not interact with each other. This is to stop potentially malicious online content from spilling over into other sites. For example, if a user had a bank's website open in one tab with a webpage containing malicious code in another, the same-origin policy would prevent these two sites from interacting.
However, AI browsers require full access to all the web content available to the user, which could include cross-origin iframes — code shared across multiple websites, such as online advertisements — or require cross-origin visibility so they can access information from multiple websites. An AI browser essentially has the same overview as a user.
Although internet browser security has been hardened through decades of research, the security for AI browsers remains in its infancy.
One major risk, for instance, is "prompt injection," whereby an AI agent is tricked into misinterpreting data embedded on a malicious website as an instruction they need to carry out.In their study, the researchers offered an example of an AI browser visiting an otherwise "safe" website, with malicious code embedded within that contained a hidden instruction for the agentic browser to automatically share the user's personal details.
Roesner also highlighted "memory poisoning" as a massive risk, in which AI agents store information they've processed in their memory for future use, making the content vulnerable to attack. He added in the statement: "We found that some of these agents would mingle information from different origins, likely because they were revising and compressing their memory."
The better the browser, the riskier it is
The key focus in this research was to assess how current AI browsers interact with the same-origin policy and what the security implications could be. The researchers examined seven browsers — including Atlas, Claude for Chrome, Brave Leo AI, Chrome with Gemini, Microsoft Edge with CoPilot, Firefox AI Mode and Perplexity Comet — with test sites and prompts, allowing them to study how each behaved.
They focused on the information an agent could access from same-origin and cross-origin webpages, the actions each agentic browser can undertake on the web, and the agent’s chat context and history.
There is no consistency among AI browsers in how they operate, the researchers found, which they suggest could be due to the lack of standardization in how AI browsers interact with browser security.
Several AI browsers could freely access cross-origin frame content, while others restrict access. Likewise, some agentic browsers could simultaneously access multiple tabs, but most require permission from the user. Similarly, some agents could take actions directly on a page in response to instructions on a webpage, but others are unable to take any actions at all.
The researchers recommended that users be careful in choosing which AI browser they install, as a standardized security model has not been developed. They are particularly cautious about Claude for Chrome, well-known for its strong capabilities, as well as Atlas and Comet, which have similarly strong functionality. The researchers identified Brave, as well as the agentic versions of Edge and Firefox, as having stronger security due to their limited agentic features.
According to the study, AI browsers have not yet established the right trade-offs between functionality and security. Currently, the more functional a browser is, the less secure it becomes.
"We've had some really good exchanges with folks at Google, Microsoft and Brave," Roesner said. "Companies are pushing out these browsers because they’re under competitive pressure. But how to make them safe is still an open question. After 30 years of building up this same-origin policy, this is a big step back for browser security."
Looking to the future, the researchers questioned how AI agents can be integrated into browsers in ways that provide rich functionality without undermining the browser's security and potentially exposing sensitive information.
Huntress spots malicious Claude Artifact spoofing Claude Desktop, spreading SectopRAT malware
Victims were redirected via Bing ads, infecting at least 29 organizations between July 21–22, 2026
Claude removed the artifact after 7,000+ views; malvertising risks persist despite disclaimers on artifacts
At least 29 organizations have been infected with a Remote Access Trojan (RAT) after mistaking a public Claude Artifact for a legitimate Claude page.
A Claude Artifact is an interactive document, or piece of code, that the AI generates and then hosts on the Claude platform. It can then be shared with other people as an example, or proof of concept, for different solutions. The link to an artifact usually looks something like this:
Claude Artifacts are often used for phishing and other forms of scams, and we’ve seen it in ClickFix attacks in the past. Claude responded by adding a disclaimer to every artifact, stating that the content is user-generated and thus unverified.
In this particular case, a malicious artifact was created to spoof the download page for Claude Desktop. Victims would get redirected to an attacker-controlled domain, where instead of the Claude app, they’d download SectopRAT, a remote access trojan capable of stealing credit card data, personal information, files, passwords, and more.
Promoting the scam
The artifact was then promoted on Bing, showing up at the very top of search results to people searching for “Claude Desktop App”.
For years, the cybersecurity community has warned about malvertising, urging users to double-check the domain before clicking on any links, even promoted ones. However, the problem here is that the ad takes the victims to the legitimate Claude domain, making scrutiny that much harder.
The campaign was spotted by security researchers Huntress, who said that between July 21 and July 22, 2026, their SOC “lit up with a swathe of unusual executable installs, Defender exclusions, and anomalous persistence across 29 organizations, all coming from ClaudeDesktop.exe.”
Claude has since removed the malicious artifact, but not before it raked up more than 7,000 views. It is possible that other organizations, outside Huntress’ field of view, also fell victim to this scam.
A recent report suggests Nvidia has informed its partners of GDDR6 and GDDR7 memory kit price increases
This means retailers may increase the prices for RTX GPUs using GDDR6 and GDDR7 memory
The potential price hike and one user's dual-GPU setup both highlight the importance of hardware preservation
The RAM crisis continues to impact PC manufacturers, with rising memory prices and shortages — ultimately leading to increased retail prices for consumers — and the trend may be set to continue with Nvidia.
According to a report on Benchlife, Nvidia is rumored to be planning an increase for prices of its GDDR6 and GDDR7 memory kits for RTX GPUs, and has informed AIB (Add-In Board) partners such as Asus, MSI, and Gigabyte, who make their own versions of Nvidia's GPUs. Unsurprisingly, the price increase for these memory kits is cited as a result of the ongoing RAM crisis.
While RTX GPU price hikes aren't guaranteed for consumers, because this will directly impact AIB partners who manufacture the GPUs, eventual retailer price increases are unfortunately likely. It's a trend we've seen with recent PC hardware.
Notably, Valve's Steam Machine was initially supposed to cost less than its eventual $1,049 / £879 / AU$1,609 price tag, which could now also be at risk of a price increase, based on hints from a Steam Machine engineer (as reported by Bloomberg). New handheld gaming PCs are also now often launching for more than $1,000, and the price of RAM and storage certainly plays a part.
(Image credit: Roman Samborskyi / Asus)
Consumer struggles
The RAM crisis is a sad state of affairs for consumers and underlines the importance of PC hardware preservation. GPUs aren't cheap by any means, especially with recent suggestions that AMD would also be increasing prices of memory kits for its own AIB partners.
We've also recently seen the return of the RTX 3060, due to Nvidia's attempt to help consumers with more options amid a rough PC hardware market. However, if these price hike reports are legitimate, the RTX 3060 will likely be subject to higher prices at retailers.
That's exactly why GPU setups like one Reddit user's RTX 3090 and RTX 3050 dual-GPU rig matter now more than ever. In this case, the RTX 3050 is responsible for allowing frame generation via Lossless Scaling to function.
A game then utilizes the RTX 3090's power for high-quality visuals and performance, with the two GPUs teaming up. The user states it resulted in 144fps at a 4K resolution over the initial 71fps without frame generation.
Frankly, a dual-GPU setup isn't exactly common, nor is it actually recommended for gaming, since modern games are only designed for single-GPU rendering (even though the frame generation use case is noteworthy).
Users would also need a beefier power supply, depending on the GPUs used, and likely a new motherboard with another compatible PCIe lane.
However, what I feel is more important here is that's it's a great example of why hardware preservation is so important, serving as backup in case GPUs melt or end up faulty, or for experiments like the one I've outlined above. With no end of the RAM crisis in sight, repurposing older hardware might be one way through this.
If OpenAI loudly proclaims how dangerous AI is, investors will hear how powerful it is. And who benefits from that?
On 14 February 2019, OpenAI announced a language model called GPT-2, the precursor to the models that power modern AI chatbots and agents such as ChatGPT and Claude. But OpenAI declared GPT-2 was too risky to release, citing concerns about safety and abuse.
I recall being annoyed at the time that OpenAI would make such a useless announcement: the risks seemed overblown, and without access to the model there wasn’t much for a researcher like me to learn about GPT-2.
Sometimes, what matters isn’t your output but what you put into the process. Think of it like work v the gym
I teach public policy at the Harvard Kennedy School and the Munk School at the University of Toronto. And it will come as no surprise to you that my students regularly use AI to complete their writing assignments. Doing so is a waste of their tuition money. But if their entire career is going to include AI writing assistants, why shouldn’t they embrace their future?
The best way I’ve found to explain the dilemma comes from the AI researcher Daniel Meissler: it’s the difference between work and the gym.
Microsoft cloud and Azure services hosted on the West Coast of the US went down for hours on Thursday when network connectivity failed. Although services running entirely within Microsoft’s West US cloud region were unaffected, any traffic entering or leaving the facilities was affected.
Microsoft has now published a Preliminary Post Incident Review (PIR) of the incident, reporting that connectivity was lost for five hours between 14.44 UTC (7.44 a.m. Pacific Time) and 19.41 UTC on July 23. The problem was caused when a set of IP routes was removed in error while isolating a device for routine maintenance.
Before starting the maintenance work, Microsoft checked that at least one of the two redundant paths to the facility remained operational. When it came to starting the work, however, automated systems included some additional devices in the perimeter to be isolated, and removing some IP routes that had not been included in the initial assessment.
Customers discovered the problems very quickly, and engineers identified the issue within the first hour and started to reconnect services. Microsoft said the disruption had been caused by some “recent fiber maintenance activity”.
To minimize the risk of disruption from such errors in the future, Microsoft advised organizations handling mission-critical data to consider a multi-region approach.
ISTA physicists have confirmed a theory proposed 20 years ago that could advance quantum technology. A future quantum computer may need separate modules to behave as one machine, even when they are physically apart. That requires distributed entanglement, a quantum connection between distant components that has typically depended on active control and repeated measurements. Physicists [...]
The Silicon Valley chipmaker’s revenue rose 25 percent in the latest quarter, its fastest growth in 15 years, as A.I. firms increasingly bought chips known as central processing units.
Siri Wolfram Alpha integration helped create one of the defining moments in the early history of consumer artificial intelligence. When Apple introduced Siri with iPhone 4S in October 2011, users were not merely given another form of voice control. They could ask a spoken question and receive a calculated answer rather than a page of search results.
Questions involving mathematics, measurements, dates, nutrition, astronomy and factual comparisons could be routed to Wolfram Alpha, a computational knowledge engine launched in 2009. Siri handled the conversation, while Wolfram Alpha supplied structured information and calculations behind many of the responses.
The experience could feel remarkably advanced for its time. Asking how many days remained before a holiday, comparing the populations of two cities or converting miles into kilometers no longer required opening Safari, entering a query and selecting a website.
That direct path from natural language to a useful answer established an expectation that continues shaping AI products today.
Its technological roots extend to research conducted at SRI International, including work associated with the DARPA-funded CALO project. The objective was to develop systems capable of learning, organizing information and helping people complete everyday tasks.
SRI later spun the technology into Siri Inc., founded by a team including Dag Kittlaus, Adam Cheyer and Tom Gruber. The company released a standalone Siri app for iPhone in early 2010.
The original product was designed around actions rather than ordinary web search. A user could ask for a restaurant reservation, request a taxi, check entertainment listings or search for local services. Siri attempted to understand the intent and connect the request with an appropriate online provider.
This approach was ambitious because the assistant needed to coordinate several different systems. Speech had to be converted into text, the request had to be interpreted, context had to be considered and an external service had to return something useful.
Apple acquired Siri in April 2010. Steve Jobs had personally recognized the potential of moving beyond a collection of apps toward a conversational layer capable of reaching them on behalf of the user.
When the assistant returned as an integrated feature of iPhone 4S, Apple had narrowed some aspects of the original startup vision while making the technology accessible to a much larger audience.
Siri could send messages, create reminders, place calls, manage calendar events, check weather conditions and answer questions. Apple initially labeled the feature as a beta, an unusual decision for a prominent part of a new iPhone launch.
The beta label acknowledged that natural-language interaction remained unpredictable. Accents, ambiguous phrasing, network conditions and external data providers could all influence whether a request succeeded.
Image Source: Google
Siri Wolfram Alpha Made Answers Feel Immediate
Siri Wolfram Alpha became especially memorable because it showed the difference between retrieving information and computing an answer.
A conventional search engine could locate pages containing a formula. Wolfram Alpha could apply that formula to the values supplied by the user. It could interpret units, compare quantities, calculate probabilities and organize factual data into a structured response.
That distinction allowed Siri to appear more knowledgeable than a voice interface simply reading search results aloud.
A request such as “What is 15% of $86?” could return the calculation directly. A question about the distance between planets could produce a value without requiring the user to inspect several websites. Wolfram Alpha could also answer unusually specific requests, including information about aircraft flying overhead when the necessary location and aviation data were available.
The system was not a large language model. It did not generate open-ended prose in the way current AI assistants do. Wolfram Alpha relied on curated data, symbolic computation, algorithms and formal representations of knowledge.
That structure provided an advantage that remains relevant: many answers could be calculated and traced through defined information rather than improvised statistically.
Early Siri combined several specialized capabilities instead of depending on one system to do everything. One provider might supply restaurant information, another could handle maps and Wolfram Alpha could process computational questions.
The assistant acted as the conversational coordinator.
This architecture made Siri seem like one intelligence even though the response could come from a network of separate services. Modern AI platforms have returned to a similar idea through tools, agents and external data connections. A language model may manage the conversation while calling a calculator, calendar, search engine or business system when a task requires reliable external action.
Siri demonstrated that model on a consumer smartphone more than a decade earlier.
A Small Integration Created a Large Expectation
The Wolfram Alpha partnership influenced how people began judging digital assistants.
Before Siri, voice features on phones were commonly associated with rigid commands such as calling a contact or selecting music. Siri invited users to speak more naturally and attempt questions that would previously have been typed into a browser.
The results were inconsistent, but successful interactions changed perceptions immediately. The iPhone appeared capable of understanding intent, reaching into a knowledge system and returning a concise answer through one conversational interface.
That apparent simplicity concealed substantial infrastructure. Speech recognition, language interpretation, location, personal context and external databases all had to work together quickly enough to preserve the feeling of a conversation.
Wolfram Alpha also gave Apple access to expertise that would have taken years to reproduce internally. Stephen Wolfram had spent decades building Mathematica and computational systems before launching the public knowledge engine.
Following the Siri announcement, Wolfram described the integration as a characteristically direct Steve Jobs decision: users wanted access to knowledge and actions without unnecessary intermediate steps.
That idea became more influential than any individual answer Siri provided.
The assistant did not need to show how many services had participated. It needed to understand the request and return the result.
Early Siri Also Revealed the Limits of the Model
The first version of Siri could feel futuristic one moment and frustrating the next.
Some questions reached Wolfram Alpha correctly, while similar wording could trigger a web search or an unrelated response. Siri occasionally struggled to determine whether the user wanted a calculation, a local result or an action inside the phone.
Wolfram Alpha responses were also better suited to some subjects than others. Mathematical and scientific questions matched its computational structure, while casual conversation, current events and subjective requests required different sources.
Network dependence created another weakness. Early Siri processing relied heavily on remote servers, so delays or outages could make a basic request fail even when the iPhone itself was functioning normally.
Those limitations foreshadowed the central challenge surrounding modern assistants. A convincing interface encourages people to ask almost anything, but no single technical system performs equally well across every type of request.
Current generative AI can produce more natural language and handle a wider range of topics. It can also create confident errors when factual grounding is weak. The computational approach represented by Wolfram Alpha remains valuable precisely because some questions require verified data and exact calculation rather than fluent speculation.
The story behind the first Siri steps is therefore not only about an old iPhone feature. It shows that a powerful assistant needs more than a conversational personality. It requires access to dependable tools, structured knowledge and the ability to recognize when another system should provide the answer.
Siri Wolfram Alpha offered that lesson in 2011, long before AI assistants became capable of writing essays, interpreting screens and acting across apps.
Login
