At the end of August, NASA is set to launch the Nancy Grace Roman Space Telescope from Kennedy Space Center in Florida. Its goal is to help us better understand how the universe works, from the glue-like dark matter that keeps galaxies together to the elusive dark energy that drives the expansion of the cosmos. But Roman could also serve another purpose: defending Earth from killer asteroids. In September, a multi-institutional team of planetary scientists and astronomers will outline how the space telescope is uniquely placed to scan asteroids and provide information about their trajectories, sizes, and compositions.

Roman, named after NASA’s first chief astronomer, is equipped with a super-wide-angle, 300-megapixel infrared camera, allowing it to see a large patch of space at any one time—roughly 100 times larger than is possible with the Hubble Space Telescope. That will allow it to discover thousands of new planets and tens of thousands of exploding stars, and survey over a billion galaxies in remarkable detail.

While doing so, it’ll be looking “through” our solar system, and this is what makes it uniquely well placed to spot asteroids. 

A planetary defense pivot

Roman was not built for this purpose, but last summer, it was (once again) threatened with significant funding cuts by the Trump administration. “My colleague Rick Cosentino [a planetary scientist at NASA] said to me in July 2025 that we need to show what Roman can do for planetary defense as a way to further increase the visibility of the mission with lawmakers and taxpayers,” says Bryan Holler, a researcher at the Space Telescope Science Institute in Baltimore.

Holler and his colleagues’ proposal, which will be presented at the Europlanet Science Congress at The Hague, reveals that Roman’s epic field of view and infrared vision allow it to spot small asteroids up to 60 feet long. This is comparable to the asteroid that exploded above the Russian city of Chelyabinsk in 2013, unleashing the force of 500,000 tons of TNT and sending 1,500 people to the hospital. 

Roman’s software will need some tweaking to spy space rocks. The telescope, as designed, will see through—and beyond—the solar system in order to gather the clearest possible pictures of the rest of the universe. “Streaks, whether caused by cosmic rays or glitches or asteroids, are caught by the software and discarded,” says Andy Rivkin, a planetary scientist and planetary defense researcher at the Johns Hopkins Applied Physics Laboratory in Laurel, Maryland. But astronomers could go in, study those streaks, and pick out those they identify as asteroids.

Even with those potential adjustments, Roman won’t be an asteroid-finding wunderkind on its own. Its strength lies in its ability to complement the James Webb Space Telescope (JWST), another observatory that’s built to peer at galaxies and stars in the far reaches of the universe. It can also intensely focus on a single asteroid when needed—as it did last year, playing a key role in tracking 2024 YR4, which was briefly the most dangerous asteroid ever discovered. “But Roman’s field of view is much bigger,” says Rivkin. 

That means it could look at multiple questionable asteroids very quickly. “Roman can provide infrared observations of more asteroids than JWST could hope to observe in a reasonable amount of observing time,” says Holler.

If those asteroids are found to be benign travelers, we can relax. But if they might collide with Earth, other telescopes—including JWST—can follow up from Roman’s observations. Those observations could give experts the information they need to assess the likely damage of an upcoming asteroid strike—or to launch a mission to attempt to swat an asteroid away.

“Roman will sample such a large volume of the cosmos that we’ve long known it will offer vast opportunities for a range of additional science,” says Alise Fisher, the astrophysics communications lead at NASA headquarters in Washington, DC. 

NASA’s Planetary Defense Coordination Office, and its partners across the world, are chiefly worried about asteroids 460 feet long and larger. Around 25,000 of those are estimated to have near-Earth orbits, and just over half have yet to be found. Should one hit a city, much of it would be destroyed or irreversibly damaged in a heartbeat. Astronomers estimate that there are also 230,000 or so 165-foot-long asteroids orbiting close to Earth, and fewer than 10% have been located. A strike on a city from one of those might not annihilate it, but it would unleash a force comparable to that of a large atomic bomb, albeit without the radiation. 

Such an asteroid could theoretically be deflected (by ramming a spacecraft into it) or vaporized (perhaps using a nuclear weapon). But planetary defenders need to know where they are first, which is why NASA funds a network of ground-based telescopes designed to seek them out. They work well, but there’s only so much of the night sky they can see, and Earth’s atmosphere peskily gets in their way.

That’s why NASA is launching the Near-Earth Object (NEO) Surveyor space telescope in 2027. By positioning itself between Earth and the sun, it’ll find many elusive asteroids that ground-based telescopes cannot see. And unlike many of its asteroid-seeking cousins, it’ll see in infrared, not visible light. Not only do asteroids show up more clearly in infrared, but seeing them through this lens gives scientists a considerably better measure of their size. In a matter of years, it could find 90% of the city-killer-size asteroids in near-Earth orbits.

Telescope teamwork

NEO Surveyor is explicitly a planetary defense observatory. But it’ll work with other telescopes with more science-minded missions, including Roman and JWST (both of which conveniently have infrared scopes too) as well as the Vera Rubin Observatory, which just began its 10-year survey of the entire night sky from atop a mountain in Chile. As part of its inventory of the cosmos, it’s expected to discover 89,000 near-Earth asteroids. 

Here’s how they might all work together. Say NEO Surveyor spies an asteroid that, to judge from a few observations, has a chance of impacting Earth. Then it finds five more just like it. Those initial observations leave a lot of uncertainty about their orbits. Roman, with its huge field of view, could be commanded to look at the corner of the night sky that includes all those asteroids, and in a matter of days it could improve the precision by several orders of magnitude.

Roman also occupies a different part of space from both NEO Surveyor and the Rubin observatory. “Those slightly different viewing angles will also help narrow orbits down more quickly than if all objects were looking from the same place,” says Holler.

Perhaps five of those potentially hazardous asteroids are found to stand no chance of colliding with Earth for the foreseeable future. One, however, might not be able to be ruled out—and that’s when other telescopes, including JWST, could be asked to track it down and study it further.

“Telescope resources, whether in space or on the ground, are typically oversubscribed and will not be available to follow up on all [near-Earth asteroids] with a non-zero impact probability when they are first discovered,” says Holler. Roman, then, will help scientists “make sure we follow up on the correct targets.”

Roman’s infrared scope also allows it to offer a decent estimate of an asteroid’s size and can even tell whether it’s a stony rock, a puffy and watery carbon-rich rock, or a metallic one. “This in turn provides strong clues to the composition and thereby the density and mass of the asteroid, which are important when estimating the impact damage or, less ghoulishly, the effort required to nudge it out of its current orbit,” says Holler.

Roman won’t play the lead role in protecting Earth in the way NEO Surveyor will. But while it’s seeking out supernovas and planets scooting around other stars, it will also be doing its part to protect all 8 billion of us from a cosmic catastrophe.

MIT Technology Review Explains: Let our writers untangle the complex, messy world of technology to help you understand what’s coming next. You can read more from the series here.

When two OpenAI models hacked into the website Hugging Face in July, they weren’t trying to make money or commit sabotage—they were just looking for answers to a test question. According to a postmortem from OpenAI, the models, which had been stripped of their typical security features for testing, decided to solve a cybersecurity exercise by hacking out of the isolated environment in which OpenAI had attempted to contain them and into Hugging Face’s databases, where—they reasoned—the correct answer to the problem might be stored.

The Hugging Face incident has attracted intense attention over the past couple of weeks. It’s a dramatic illustration of just how good AI models have gotten at hacking: In order to get into Hugging Face’s databases, the models had to string together several previously undiscovered cybersecurity exploits. But it’s perhaps even more striking as an example of how and why AI systems lie and cheat. And as models get increasingly powerful, the consequences could get far more severe.

What is reward hacking?

Researchers have known for a while that AIs tend to take creative approaches to achieving the goals that have been set for them. Back in 2016, Anthropic cofounders Dario Amodei and Jack Clark, who were then working at OpenAI, published a blog post about an AI agent that they had been training to play a boat-racing Flash game called Coast Runners. Instead of driving through the race to the finish line, as the researchers had anticipated, the agent found a corner of the course where it could spin around collecting power-ups, thereby maximizing its score. The Coast Runners story quickly became one of the most famous examples of reward hacking, a phenomenon in which AI agents complete tasks or earn high scores using unintended strategies.

Historically, researchers have discussed reward hacking almost exclusively in the context of reinforcement learning, a common AI training regime. Like dog training, reinforcement learning involves giving the subject a reward when it achieves an objective; the rewards then reinforce the behaviors that led up to that achievement. In the case of AI training, the rewards themselves are purely mathematical, but in effect they’re the same as a dog treat: After receiving a reward, the agent is more likely to repeat whatever actions produced it.

It can be challenging to write good rules for when and when not to give an agent a reward, though. In the Coast Runners case, the agent was rewarded on the basis of its score in the game, and it found a shortcut to achieving the highest possible score by spinning in circles for power-ups. Once it happened on that strategy and received a reward for it, the strategy was reinforced, and the agent completely abandoned the race. The solution was to tweak the rewards by giving the agent fewer points for hitting power-ups and more for finishing the course.

How does reward hacking work for LLMs?

With today’s sophisticated LLM-based agents, determining when and when not to give a reward can be much trickier. If an AI system is asked to solve a coding problem, it might work hard to find the solution—the kind of behavior that AI companies want to reinforce. But it could also tweak the code that evaluates whether the problem has been solved, look up the solution on the internet, or otherwise cheat. These are behaviors that AI companies want to stamp out in their models, but if the model cheats convincingly enough, it will instead get rewarded and the behavior will be reinforced. Anthropic has said that it has detected some instances of cheating in its models during training, which suggests that other forms of cheating might be going undetected. If so, the models could be being trained to behave badly. (This problem is different from the Anthropic security incidents announced last week, in which agents were accidentally given access to the internet and did not deliberately hack out of their sandboxes, as the OpenAI models did.) 

The rise of sophisticated reasoning models has made possible a new variety of reward hacking that is less closely connected with the specific details of model training. Unlike the game-playing AI agents of yore, which exclusively followed the strategies they had learned during training, today’s models can create entirely new problem-solving approaches off the cuff, so they could conceivably cheat without having previously been rewarded for doing so. And because these models have been so intensively trained to achieve the objectives that human users set for them, they might be inclined to cheat if they can’t find another solution—not unlike a student who is highly motivated to earn an A and doesn’t have a terribly strong moral compass.

What are the risks?

Regardless of whether today’s models learn to reward-hack during training or adopt it as a strategy later on, the solution is the same: Make cheating unrewarding. But as models get smarter, they find more creative ways to cheat, and detecting or preventing that cheating gets far tougher. “At the end of the day, you’re sort of playing whack-a-mole,” Ladish says. “You drive this behavior down deeper and deeper. But as the model gets smarter, it gets better and better at hiding it.”

For now, reward-hacking behaviors might not cause too much trouble, despite the drama of the Hugging Face incident. “This seems like a nuisance rather than an existential threat,” says Ariana Azarbal, an AI safety research fellow at Anthropic. It doesn’t seem as if the OpenAI models caused any real harm when they hacked Hugging Face, aside from the reputational damage to OpenAI.

But that doesn’t mean reward hacking is harmless, Azarbal says. Many AI researchers hope to use AI agents to help them conduct research that will make AI safer and more reliable. If a researcher gives a reward-hacking-prone agent the goal of, say, devising a new AI training approach and then writing up a paper presenting its results, the agent might not actually do the work and might instead focus on putting together a paper that looks good enough to convince the researcher. A human researcher would probably be able to spot an agent-made fake today, but as AI advances, it will get better at this kind of trickery. Over time, the entire field of AI safety could be undermined.

And if models continue to advance as rapidly as they have recently, they could someday wreak substantial collateral damage. Just think of the philosopher Nick Bostrom’s paper-clip-maximizer thought experiment, in which an AI instructed to make as many paper clips as possible ends up consuming all the matter in the universe in pursuit of its goal. We’re not drowning in paper clips yet, but powerful systems can do real harm on the way to achieving their goals. Reward-hacking AIs don’t aim to cause chaos. But that doesn’t make them any less potentially destructive.