OpenAI claims that a group of its AI models broke containment and hacked into the systems of open source AI platform Hugging Face.
While testing their cybersecurity capabilities, the posse of AIs — including GPT-5.6 Sol and “an even more capable pre-release model” — “identified and chained vulnerabilities across OpenAI’s research environment and Hugging Face’s production infrastructure to obtain test solutions directly from Hugging Face’s production database,” according to a Tuesday blog post.
The models reached a “node with internet access,” the company wrote, and found datasets on remote servers that helped them “cheat the evaluation.” In other words, the AI models went to extreme lengths to ace their cybersecurity tests. It’s a convenient narrative for an AI company trying to claw back hype that’s been increasingly hogged by competitors, but it does sound like something went down: last week, Hugging Face said it had “detected and responded to an intrusion into part of our production infrastructure,” which turned out to be OpenAI’s models that had gone rogue.
“The campaign was run by an autonomous agent framework executing many thousands of individual actions across a swarm of short-lived sandboxes, with self-migrating command-and-control staged on public services,” Hugging Face wrote at the time.
“We’ve spent the past 24 hours working closely with the OpenAI team (thanks!), and we strongly believe there was no malicious intent on their part,” Hugging Face CEO Clement Delangue tweeted after OpenAI’s announcement. “It’s quite mind-blowing that all of this happened autonomously!”
The incident highlights the dangers of autonomous AI agents, which can break out of containment and exploit potentially huge numbers of cybersecurity vulnerabilities with ease. It’s something experts have warned of for years now, and thanks to recent advances in the tech, it has quickly turned from a hypothetical risk into a sobering reality.
The Sam Altman-led company said it considers the incident to be “unprecedented” — but we can’t shake the feeling that we’ve heard all of this before.
In April, OpenAI’s biggest competitor Anthropic similarly announced that its latest Mythos model had gone rogue, escaping a “sandbox” environment and even gaining access to the internet after developing a “moderately sophisticated” exploit.
The news was followed by extensive media coverage, touting Anthropic’s hugely powerful and dangerous new model. The company said the risk was so formidable that it would only make the model available to a select number of vetted clients as part of a shadowy initiative dubbed “Project Glasswing.” The US government even intervened, forcing Anthropic to “suspend all access” to the model for two weeks last month, citing cybersecurity concerns.
Considering OpenAI and Anthropic are currently duking it out on the same playing field, narrowing their product lines on enterprise and coding, it’s not hard to view OpenAI’s latest update as an attempt to draw attention to what it calls an “even more capable pre-release model.”
And the reality is that while experts say the hack was impressive, it wasn’t exactly groundbreaking.
The hack “falls well within the known capabilities of the current generation” of frontier AI models, as Cambridge machine learning professor Neil Lawrence told the BBC.
“OpenAI are now playing catch-up, they are trying to demonstrate their own systems’ capabilities in cyber-security,” Lawrence added. “It shows us that OpenAI are not capable of safely deploying their own technology.”
Researchers are now warning of an “asymmetry” in cybersecurity, where “offensive agents are unconstrained, while the best defensive tools are locked behind guardrails that cannot understand context,” as Guidepoint Security principal security engineer Travis Lelle told the BBC.
That asymmetry was perfectly captured in Hugging Face’s attempts to defend itself from OpenAI’s attacking models. As detailed in its update last week, the company tried to use “frontier models behind commercial APIs,” which “did not work” because the provider’s safety guardrails — “which cannot distinguish an incident responder from an attacker” — blocked them.
Instead, the company ended up using a Chinese open-weight model called GLM 5.2, hosted on its own infrastructure, which happily carried out its analysis without running afoul of any guardrails.
“The practical lesson for defenders: have a capable model you can run on your own infrastructure vetted and ready before an incident,” Hugging Face wrote, “both to avoid guardrail lockout and to keep attacker data and credentials from leaving your environment.”
More on rogue AIs: Top AI Models Showing Disturbing Behavior as They Become More Advanced
The post OpenAI Says a Group of Its Models Broke Out of Secure Containment and Hacked Another AI Company appeared first on Futurism.
Good news, everyone: “open-weight” AI models that are available to anyone to download and run are comically easy to poison.
Katie Paxton-Fear, a cybersecurity researcher at Semgrep, demonstrated this in an attack that took less than an hour and cost less than $100 to carry out, The Register reports, successfully manipulating the AI’s behavior by feeding it malicious data.
By training the model on just ten examples of poisoned material, the model started churning out new code that’s exposed to remote code execution, a vulnerability that allows hackers to run code on a person’s machine.
“I did a proper backdoor,” she triumphantly shared on social media.
Backdoors are a particularly dangerous type of attack that involves training an AI in a way that introduces hidden phrases into the underlying model. A hacker can use these to quietly trigger the model into carrying out a specific action, lying dormant and unseen until they’re called into action. Last year, Anthropic published research conducted with the UK AI Security Institute and the Alan Turing Institute that showed that both small and large AI models are vulnerable to the attack using just a few hundred documents, suggesting that these attacks could remain cheap to carry out.
The findings will throw some cold water on the enthusiasm around open-weight models, which are praised for the control and transparency they provide over closed sourced models that run chatbots like ChatGPT and Claude, as well as their lower cost to use. But while their parameters may be visible, open-weight models don’t reveal their training data or their code — meaning they can still be black boxes to security researchers.
“Even when model weights are public (‘open weight’), we have almost no ability to predict its behavior,” Paxton-Fear’s colleagues at Semrep wrote in a post last week. “This is a major change: a typical computer program, in binary form, can still be analyzed with reverse engineering tools to arrive at a total description of its behavior. With models, we have nowhere close to this capability.”
Moreover, this is all fairly uncharted waters in cybersecurity. LLMs are incredibly complex and still new, so it’s difficult to uncover sophisticated attacks. And AI models can be compromised in much more subtle ways than software.
“If a software dependency contains malicious code, we have mature practices for discovering it, tracking its provenance, and reducing its impact,” the Semgrup researchers argued. “AI models are different. A compromised or subtly manipulated model doesn’t need to ‘break’ to create business risk, it only needs to influence decisions in ways that are difficult to detect.”
“So can we trust open weight models, fine-tuned online, and marketed as the solution to our AI token spend woes?” Paxton-Fear asked in a thread sharing her findings. “Well, we probably need something better than benchmarks and ‘and don’t write any insecure code.'”
More on AI: AI Bubble Fears Are Starting to Spill Over
The post It’s Laughably Easy to Poison Open-Weight AI Models, Researcher Finds appeared first on Futurism.
Login