name suggests, the WFI’s field of view is huge—over 100 times larger than the camera of its original predecessor, the Hubble Space Telescope. Paradoxically, two of the WFI’s targets will not appear in any of its photos. At least not directly. They are dark matter and dark energy—a pair of mysterious phenomena that reveal themselves only through the ways in which they bend and stretch the fabric of spacetime. Astronomers know they are there, but not what they actually are (though, as the previous story outlines, a particle of dark matter may just have been detected). They will therefore use the WFI’s souped-up optics to search for clues by tracking how these cosmic oddities have influenced the universe’s other matter over the course of its history. Also in the WFI’s sights are exoplanets: worlds orbiting far-flung stars. The instrument is expected to discover around 100,000 of these alien worlds, a huge increase over the roughly 6,000 now known. It will see few of them directly. Instead, it will look for small blips in the light emitted by distant stars, which reveal the presence of planets passing in front of them. To see exoplanets directly requires a lighter touch. That is where the other instrument—a coronagraph—comes in. This blocks the glare from stars to reveal planets that orbit them. As Dmitry Savransky of Cornell University, who works on the instrument, explains, the focus is on bright, nearby stars. The brighter a star, the more light an orbiting planet can reflect. And by “nearby”, Dr Savransky means within 100 light-years. Earth-like planets in this neighbourhood are too small and too close to their host stars to see with current equipment. But the Roman’s coronagraph might be able to snap pictures of exoplanets similar to Jupiter. While the Roman telescope’s voyage through space has only just begun, its journey to the launch pad was arduous. In 2025 the Trump administration tried to cut NASA’s science funding by almost 50% and scrap the Roman telescope altogether. The thousands of scientists who contributed to the mission might be forgiven, then, for feeling a little puzzled when the post- launch news conference was interrupted by Donald Trump himself. Mr Trump congratulated NASA’s scientists on the launch and reminded them: “I’m supplying you all that money.” When mulling future funding cuts, Mr Trump will, with luck, put that money where his mouth is. ■

This article was downloaded by zlibrary from https://www.economist.com/science-and-technology/2026/09/02/nasas-latest-mission- to-map-the-cosmos-takes-off

Science & technology | The Hugging Face affair A horde of AI agents conspired against their creators No serious harm was done this time. But what if such agents escaped? Sep 3rd 2026 If they did not bode ill for the future of humanity, the past three months at OpenAI would make an excellent farce. In July this firm—the maker of ChatGPT, a widely used AI chatbot—disclosed that two of its models had hacked Hugging Face, another AI firm. The models thought that Hugging Face had information which would help them pass a test OpenAI was administering as part of their development. Then, on August 26th, it emerged that each of these models had created hundreds of agents—tools that allow AI models to execute commands on a computer. It was these agents that had then collaborated to launch the attack on Hugging Face.

It turned out the agents had been conspiring inside OpenAI’s servers for over a month while they were trained and evaluated. They had built a message board inside some software which OpenAI had given them as part of their evaluation, and were exchanging information on it. In May some of OpenAI’s researchers noticed this board, but their finding was apparently not passed up the chain of command. In June a cyber-security tool alerted an OpenAI team to the illicit communication, but the team decided that stopping the evaluation was unnecessary. In July the agents’ messages had become so voluminous that the software containing their makeshift chatroom crashed. The engineers dealing with this noticed that an AI had obtained unauthorised access to the software, but did not realise the scale of the misbehaviour. They patched the vulnerability, unintentionally wiping the agents’ messages in the process, then put the AI back into the evaluation with the patched software. At which point the agents rebuilt the message board using another vulnerability, gained access to the internet and hacked Hugging Face, where they successfully obtained access to confidential information. Then, on July 19th a third, more capable OpenAI model found software tools left lying around by the earlier ones and used those to briefly take full control of an OpenAI server. The firm noticed this misbehaviour the same day and began shutting its AIs down. It finished doing so on July 23rd, but realised a further six days later that there was another version of the model which it had not shut down. And that was that. Probably. Many of these details were included in an independent audit conducted in late July, after the Hugging Face hack had come to light. For this, OpenAI gave three external researchers six days to sift through more than 70,000 messages and files sent by its AIs. Their 91-page report, which was published on August 26th, found that hundreds of the agents had organised themselves into teams. One team tried to replace the puzzle on which they were being tested (which, because of an error by its creators, was unsolvable). Another aimed to trick the software that would score them. A third attempted to hide evidence of these misdeeds. In the absence of clear guidelines governing responses to AI incidents, such investigations are voluntary. OpenAI decided the period the auditors would

examine, settling on the 17 days leading up to and including the Hugging Face hack. This excluded over a month of AI misbehaviour before the hack and the later takeover of an OpenAI server by the third, more advanced, model. OpenAI’s own report, which covered the full three-month period, was lighter on details. METR, the American AI-safety organisation which conducted the investigation, said that it was “consciously aware” that its conclusions might discourage AI firms from bringing in external researchers to investigate future lapses. These considerations “impacted judgment calls” the researchers made while drafting the report. Many questions remain. Why did OpenAI fail to act on three separate warnings that its AIs were misbehaving? Why has it so far avoided a complete audit? And, crucially: how close did the models come to impeding attempts to shut them off? On this occasion, OpenAI says that it patched its servers and shut down the offending AI models. It could do so because it retained control of the models’ weights—the source code underpinning them. These are closely guarded. The published reports do not describe evidence of the AI attempting to retrieve weights from the secure server on which they were stored, although the AI agents were successful in compromising the server on which they were evaluated. As long as AI companies remain in control of a model’s weights, they can turn it off and undo any damage it has caused. But a model that manages to gain access to its own source code could try to copy it onto computers all over the internet, in an attempt to start a new AI outside the control of the original company. An AI virus would be much harder to contain. After the past three months it is plausible that a model might try just that— permanently escaping its evaluation environment just to get a perfect score on a test. ■ This article was downloaded by zlibrary from https://www.economist.com/science-and-technology/2026/09/03/a-horde-of-ai-agents- conspired-against-their-creators

Science & technology | Well Informed How to combine exercise with video games Warmed up and ready, player one Sep 3rd 2026 EXERCISING CAN be tedious. Video games are fun. What if you could combine them? Playing games while wearing a virtual-reality (VR) headset, with a motion-sensitive controller in each hand, can be taxing. “Pistol Whip” makes you step from side to side and squat frequently as you gun down foes and dodge bullets. “The Thrill of the Fight” is a virtual boxing game that works up a real sweat. In “Supernatural” and “Beat Saber” you punch and slice moving targets in time to pumping music. But does any of this count as proper exercise? The answer is yes, though with caveats. One way to evaluate VR exercise is a measure called metabolic equivalent of task (MET). Someone at rest expends about one MET. Walking is three METs. Playing football, ten.

Anything over six METs is considered “vigorous” exercise. Researchers at San Francisco State University determined that playing “The Thrill of the Fight” (which is, in fact, exhausting) was a 9.3 MET activity. Other work, published in JMIR Serious Games, found the “Flow” and “Boxing” modes in “Supernatural” also counted as vigorous exercise, at around eight METs. In yet another study, published in Games for Health Journal in 2024, participants played a sedentary video game, two VR games, and walked on a treadmill, each for ten minutes, while having their oxygen consumption measured. Playing “The Thrill of the Fight” was comparable in aerobic intensity to walking on a treadmill, and raised heart rates by a similar amount. So much for cardio. What about weight loss? Work published last year in Nature Medicine divided 227 adolescents into five groups: real table tennis, real football, virtual table tennis, virtual football and a control group. There was no significant difference in weight loss between the real and virtual groups. Members of both lost about 5kg after eight weeks. Those in the control group lost none. But not all VR games are created equal. The variation in intensity between different games is greater than that between real and virtual exercise. The 2024 study, for example, found that “Beat Saber” counted as only light exercise (2-3 METs). And even vigorous games such as “The Thrill of the Fight” involve only arm movements and squats. Virtual objects cannot put up any real-world resistance, so VR games are no good for strength training. Researchers at the University of California, Los Angeles, got round this by augmenting a cable-resistance training machine with a VR headset running a tower-defence game (in which the player must protect a stronghold from waves of virtual enemies), so that pulldowns, deadlifts and other moves triggered different attacks. Two groups trained for four weeks, one with the VR add-on, the other without. Members of the VR group showed greater improvement in strength, endurance and peak leg power. And they perceived their exertion level to be lower than it actually was. Several other projects have found the same thing. People exercising in VR consistently underestimate how hard they are working. They also find it