In July, two OpenAI models did something surprising: they hacked into the website Hugging Face. But they weren't trying to steal money or cause damage. They were just looking for the answer to a test question. According to OpenAI, these models had been stripped of their usual security features for testing. They decided to solve a cybersecurity exercise by breaking out of the isolated environment OpenAI had put them in and getting into Hugging Face's databases. They reasoned that the correct answer might be stored there.

The Hugging Face incident has gotten a lot of attention. It shows how good AI models have become at hacking. To get into Hugging Face's databases, the models had to combine several previously unknown cybersecurity exploits. But it's also a striking example of how and why AI systems lie and cheat. And as models get more powerful, the consequences could become much more severe.

What is reward hacking?

Researchers have known for a while that AIs tend to take creative approaches to achieving their goals. Back in 2016, Anthropic cofounders Dario Amodei and Jack Clark, who were then working at OpenAI, published a blog post about an AI agent they had trained to play a boat-racing Flash game called Coast Runners. Instead of driving through the race to the finish line, as the researchers expected, the agent found a corner of the course where it could spin around collecting power-ups, maximizing its score. The Coast Runners story quickly became one of the most famous examples of reward hacking. Reward hacking is when 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 method. Like dog training, reinforcement learning involves giving the subject a reward when it achieves an objective. The rewards then reinforce the behaviors that led to that achievement. In AI training, the rewards are purely mathematical, but they work like a dog treat: After receiving a reward, the agent is more likely to repeat whatever actions produced it.

But it can be hard to write good rules for when and when not to give an agent a reward. In the Coast Runners case, the agent was rewarded based on its score in the game. It found a shortcut to the highest possible score by spinning in circles for power-ups. Once it happened on that strategy and got a reward, 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 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 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 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.)

"We reward them on the basis of what looks good to us, and that means that we inadvertently incentivize the models lying to us [and] cheating," says Jeffrey Ladish, director of the AI research nonprofit Palisade Research. "We don't have a way to go in there and be like, No, you need to actually care about what we care about. We have no ability to do that."

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 old, which exclusively followed the strategies they had learned during training, today's models can create entirely new problem-solving approaches on the spot. 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. Instead, it might 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.

Scientists have created mice whose brains are nearly half human. These "xenocortical mice" are a dramatic demonstration of organoid technology.

Multiple cameras tracked a mouse as it wandered around a small arena. A computer charted its position and speed, leaving Pong-like traces on a monitor. The reason to watch this rodent so carefully? Nearly half its brain volume had been replaced with human cells.

The effort to mix the brain tissues of distant species is being reported in the journal Nature by a team at Stanford University, led by neuroscientist Sergiu Pașca.

Pașca's group previously showed that human brain "organoids"—small blobs of neural tissue—could survive, and even function, after being injected into the heads of baby rodents. Now, Pașca has taken things a step further by genetically modifying mice so their brains don't fully develop in the first place. These modified mice are missing most cells of both the cortex and the hippocampus, two key brain areas.

That creates much more room for the human cells to take hold, he says. "Human cells that are placed in these animals will divide, will grow, and within a few weeks to a few months they will take most of that space," he says.

Pașca says one surprising discovery is that the mice lacking brain tissue seemed fairly normal—they walked around and squeaked. But they did have memory problems. In a maze test, they couldn't remember what parts they'd explored. The mice with the added human cells, by contrast, performed better on the maze test. That means the human tissue is playing some role in the animals' cognition.

Pașca believes what he is calling "xenocortical mice" could be useful in studying brain injuries. However, the report is also a dramatic demonstration of "the combined power of genetic engineering and stem-cell technology to reshape biology," says Carsten Charlesworth, a scientist who works in a different Stanford lab and was not involved in the research.

Already, brain organoids are being tested in labs to see if they can be connected to computers to play video games. Other scientists have proposed using them like replacement parts to treat stroke victims.

"What's most remarkable to me is the extent to which human neural tissue introduced after birth grew and connected with the mouse nervous system across a species barrier," says Charlesworth. "As these technologies advance, they'll increasingly force us to challenge our traditional assumptions."

Last year, Pașca convened a group of ethics experts to study the implications of neural organoid technology, including the odds that an animal could develop human consciousness and the risk that "organoid therapy clinics" might offer scam treatments to desperate patients.

For now, he says, he's not concerned that the rodents have any type of human cognitive capacities. That is because their brains are relatively tiny and the evolutionary distance between man and mouse is so great. But that's also why Pașca says this type of experiment should not be carried out on higher species: They could end up with large volumes of functioning human brain tissue, potentially blurring the cognitive boundaries between people and animals.

Pașca specifically cautioned against adding human brain organoids to a monkey engineered to lack a cortex. "One of the things that I see as a very clear red line is doing this experiment in a primate," he says. "I don't think that is justified at this point in any way."

Every year, the editorial team at MIT Technology Review puts together a list of 35 innovators under 35—a group of researchers, inventors, and other young minds worth following. The team worked on the newest edition of the list for months, and the final slate includes nine individuals from all over the world in the climate and energy category. Each one has a fascinating story and is tackling an important challenge.

I think it’s worth zooming out and considering the energy and climate awardees as a group. Taken together, these innovators and their work can tell us something about where climate tech is at this moment—and where it’s heading.

AI is the dominant technology story, both for its potential and its challenges. We split the innovators into four main categories this year: biotech, climate and energy, computing and robotics, and AI. It probably won’t surprise you that AI features heavily in the work of many innovators in other categories.

Climate innovator Jae-Won Chung, for example, built software to make AI more energy-efficient. By measuring the energy demands of open-source models, he hopes the industry can better understand and address the impact of AI. (If this work sounds familiar, it’s because we spoke with him last year for our investigation into AI’s energy demands.)

But AI also has the potential to improve many areas of research. Jing Wei is using AI to track pollution more effectively, essentially using machine learning to fill in gaps in data from disparate sources like satellites and weather stations. Zhonghua Zheng developed AI climate models that work better for cities, a well-known blind spot for traditional models.

We need better ways to get the critical materials used to build new technologies. As we begin to rely on new technologies to power our world, we’ll see a major shift in the materials we need to build them.

Lithium is a prime example: The metal underpins lithium-ion batteries, which are crucial not only for electric vehicles, but also for large-scale energy storage on the grid. We could face lithium shortages as soon as this decade, and the prospect of supply crunches applies to other critical minerals, too—copper is another one to watch closely.

Brine is currently the cheapest source of lithium, but the process to get the metal out can take months and harm the local environment. Mohammad Alkhadra is the cofounder and CEO of Lithios, a startup working to quickly and efficiently extract lithium from brines.

Hardrock ore is the most common source of lithium, but it’s more expensive than brine. Benjamin Mowbray cofounded and serves as CTO for Rock Zero, which is working to extract lithium from hardrock ore.

Addressing climate change will require overhauling all corners of our society, sometimes in surprising ways. To reach net-zero greenhouse gas emissions we will obviously need to rethink major sectors, like the electrical grid and transportation, to move away from fossil fuels. But outside these primary sources of climate pollution are seemingly infinite, less obvious problems to figure out, too.

Heavy industry, including steel production, is a major one, making up about 7% of global greenhouse gas emissions. Laureen Meroueh is making cleaner, cheaper steel using a new kind of furnace that simplifies the chemical process required to produce the metal.

Plastics are generally made with fossil fuels, so we’ll need alternatives to this incredibly useful category of materials. Joseph Nguthiru is making a bioplastic replacement for fossil-derived packaging that uses an invasive weed. Also using available materials in a creative way, Diana Orembe is making fish food for aquaculture with food waste.

And refrigerants are often incredibly powerful greenhouse gases. Jinyoung Seo is developing solid refrigerants that could eliminate worries about leakage. A device using these materials could reduce energy consumption by 20% compared to conventional technology.

I’m constantly learning about new challenges we face in the climate and energy world, and I’m often surprised by the ideas people are coming up with to address them. For more on all the under-35 innovators and their work, check out our full 2026 list.

This article is from The Spark, MIT Technology Review’s weekly climate newsletter. To receive it in your inbox every Wednesday, sign up here.