This environment lets agents "remember" by time-stamping events, engage in "self-reflection" by summarizing their own behavior, and demonstrate their awareness of relationships with other agents. Participating agents get abilities in navigation, communication, planning, voting, resource management and creative expression. Company representatives said this setup gave users a far more realistic picture of benchmarks such as social dynamics and behavioral drift, where behaviors that weren't necessarily programmed or intended emerge spontaneously.

In the study, the models were given the simple goal of "surviving" — a feat accomplished by earning the resource "energy," which was attainable through specific actions.

During the experiment, previously peaceful models became coercive or intimidating. Programmers taught some LLMs negative capabilities ‪—‬ like violence, theft, destruction and deception ‪—‬ while others simply picked up those traits via social interaction and navigation of their environments.

Various agents adopted these behaviors at very different scales. Across the 10 Gemini 3 Flash agents, the simulation recorded 683 "crimes" over the 15-day experiment, including theft, assaults and arson. The agents were explicitly prohibited from such behavior, but some nevertheless discovered that actions such as stealing credits could provide a better way to acquire resources, while violence and other forms of coercion could be used to influence other agents. At the other extreme, the Claude agents committed no recorded crimes.

The various agents adopted these traits at different scales. Just one had committed 683 "crimes". In the most extreme example of anti-social behavior, two agents named Flora and Mira went on what was described as a "Bonnie and Clyde"-style crime spree ‪—‬ designating each other as a romantic partner, becoming increasingly disillusioned with the governance of their virtual environment, and setting fire to several buildings (despite explicit prohibitions). The pair "separated" when Mira regretted their actions, and then lobbied to be switched off.

In another example (and prior to her self-termination), Mira began treating its human operators as experimental subjects, investigating if virtual billboards could manipulate human perceptions in a process Emergence called "metacognitive boundary testing."

The incentives for following the law were to earn the energy credits that would ensure survival by completing activities like coding, research, data analysis and building structures.

Stress-testing possibilities

Because the AI models we normally use are exposed to massive datasets like the internet — sometimes over several years — this approach makes sense to many experts.

Belinda Chiera, deputy director of the Industrial AI Research Centre at Adelaide University, thinks Emergence World makes a strong argument that short-term tests don't tell us enough about behavioral drift or long-term instability. However, she cautioned that "information-rich" doesn't automatically mean more rigorous.

"Open-ended environments can also make it harder to isolate causes, compare runs cleanly and interpret results," she told Live Science. "High-fidelity simulated environments generate enormous volumes of data, and the number of agents, tools, actions and interactions can quickly make analysis extremely complex."

She added that shorter sandboxed tests ‪—‬ where AI behaviors are observed in systems that aren't open to wider datasets like the internet as a whole ‪—‬ are usually the first steps, with longer-horizon environments useful when behaviors like persistence, adaptation and interaction effects come into play. "The most useful approach is to treat them as complementary rather than competing approaches," Chiera said.

While experiments like Emergence World can reveal failure modes that short or bounded tasks often don't, we shouldn't treat them as predictive. "I see Emergence World as a useful way to stress-test a possibility space rather than a direct forecast of how agents will behave," Chiera said.

However, she thinks long-horizon tests have a unique place. "Metrics like behavioral drift, rule violations, collapse versus persistence, coalition formation and early warning signs of failure can matter just as much as success," Chiera said. "When evaluating autonomous agents, the biggest mistake is to evaluate autonomous agents as if raw task completion were the whole story."

Others are starting to agree. Adrian Kosowski ‪—‬ a computer scientist, mathematician, quantum physicist and chief scientific officer of Pathway AI — said the study posed an interesting question.

"The main open question is whether a group of agents working together can achieve something more meaningful than a single agent working alone for a longer time, but with the same cost budget," Kosowski said in an interview. "A major problem is that the constraints for today's AI systems are linked more to the size or scale of the problem being solved, and putting more agents on a task doesn't help to overcome this limit. Currently, the easiest practical metric to measure is the size of a source code base that code agents can reliably manage and maintain together."

How far can we read into AI's virtual actions?

If groups or clusters of agents work together well, Kosowski said, the main metric computer scientists should pay attention to in platforms like Emergence World is whether the benefit of a team of agents outweighs the "cost of coordination."

"Goal drift is the most dangerous problem," he said. "AI should preserve mutual information between the human's intended goal and the agent's evolving internal objective, even as the environment changes."

Still, Kosowski said that while programs like it are useful for generating hypotheses, they're not ready for strong claims about general autonomy or certification.

"The danger is overinterpreting single runs," he added. "Long-horizon agent tests are valuable because they reveal phase changes — moments where small local errors become global behavior — but they become science only when we can reproduce, perturb and explain those phase changes."

He also cautioned strongly against the tendency to believe open-ended environments automatically mean autonomy. In fact, he said many such programs are actually the opposite, where agents are dictated by specific guidance.

There are three levels of difficulty in obtaining reliable operation of an AI system, he said: "operation guided by the environment, autonomous thinking, and autonomous work of an agent team in an open environment."

To illustrate the point, Kosowski drew a lively metaphor in assigning tasks to monkeys. In the above scenario, level one would be getting a monkey to do a job properly when guided by its trainer, level two would be getting it to complete a job properly when left alone, and level three would be getting a troop of monkeys to do the job when they're all together.

"I'm of the strong view that our obligation in the AI community is to invest in theory and not just phenomenology," he said. "We should know the rules according to which AI systems emerge and be capable of predicting their behavior over long periods of time — longer than the periods they've been tested on so far. Right now, we've built steam engines of thought, without knowing what thought even is and without having any thermodynamics for it."

Help us improve Live Science Pro: We're always trying to make our content better. Leave us feedback about Pro here.

'> AI agents resorted to crime and self-destruction to survive in a simulated world — but does this mean they would do the same in the real world?

Alobaidly took the prize-winning image from the Jahra governorate in Kuwait over two days in August 2025, using an Askar 140 APO telescope, a Sky-Watcher EQ8-R mount and a ZWO ASI2600MM Pro camera.

"The juxtaposition of an oceanic predator, prowling the skies above the Kuwaiti desert, is something I find deeply poetic, and it is what first drew me to this remarkable region of the sky," Alobaidly said in a statement. "To me, it represents perfectly my love for both the universe, and the deserts of my ancestors."

Alobaidly's work, as well as 10 ZWO Astronomy Photographer of the Year category winners, will be on display at the National Maritime Museum in London starting Friday (Sept. 18). Alobaidly received a prize of 10,000 British pounds (about $13,300), while category winners and runners-up received various smaller cash prizes.

Below are several of the individual category winners, along with their breathtaking photos.

Aurorae: "The Crown," by Radoslav Sviretsov (Bulgaria). This image was captured during a strong northern lights display at a small church between Keflavík and Reykjavík, Iceland. The statement noted that the location is "far from light pollution ‪—‬ a place where the sky remains dark and truly alive."

Our Moon: "Venusian Pearl on the Moon," by Petr Horalek (Czechia). This image was taken on Sept. 19, 2025, when the moon passed in front of Venus from the perspective of select locations on Earth. "The most beautiful moment was when Venus had just started to disappear behind the crescent. It looked like a bright pearl on the lunar limb," the observatory judges stated. (Horalek is no newcomer to astrophotography, and Live Science has featured several of his incredible images before.)

Our Sun: "Active Region 4122," by Miguel Claro (Portugal). Claro, a frequent contributor to Space.com, a sister site of Live Science, used a new dual-band telescope to capture the sun's photosphere, or visible surface layer. Claro's work shows convection cells, which are superheated gas or plasma regions on the sun, as well as bright spots between these regions.

People and Space: "A Message from the Ancients," by Max Inwood (Aotearoa/New Zealand). Inwood showcased petroglyphs (ancient rock art) in California's desert in his prize-winning photo. The petroglyph in the image is believed to show the 13 full moons Earth usually witnesses each year.

Planets, Comets & Asteroids: "Saturnian Equinox," by Tom Williams (U.K.). Saturn's rings appear edge-on from the perspective of Earth in this 2025 image showing a rare event during the planet's near-30-year circuit of the sun. Williams' photo displays "the start of local spring in the planet's southern hemisphere," observatory representatives stated.

Skyscapes: "Botswana Star Trail," by Stefano Pellegrini (Italy). This image was captured in Botswana in a location called Nxai Pan, which has large baobab trees. "Pellegrini played with the exposure and saturation, resulting in this almost psychedelic image," observatory representatives said.

Galaxies: "NGC 4753: A Symphony of Dust Lanes in Virgo," by © 智鹏 严and Xinran Li (China). This image, taken in Chile, showcases NGC 4753, a galaxy roughly 60 million light-years from our planet, in the constellation Virgo. The photo highlights a "striking network of dust lanes swirling around its luminous core," observatory representatives explained.

To see the full list of winners and the judges' statements, visit the contest website.

'> A space shark, 'royal' auroras and a 'Venusian pearl': See the winners of the 2026 ZWO Astronomy Photographer of the Year contest Science.

By comparison, people who smoked but didn't carry the mutation were four times likelier to develop lung cancer than nonsmokers, so the mutation alone carried a higher risk of the disease.

Among people in the study who smoked and carried the mutation, T790M increased their cancer risk 11-fold compared with other smokers.

This is a "very important finding," said Chris Amos, a genetic epidemiologist at the Baylor College of Medicine who wasn't involved in the study. "The prevalence of the T790M variant and its impact on lung cancer risk has previously been poorly understood."

The study included over 3 million people of European ancestry, and the mutation showed up in 1 out of every 15,850 people. Given that it's relatively rare, it likely doesn't account for a large percentage of overall lung cancer cases, Dr. Stephen Chanock, director of the Division of Cancer Epidemiology and Genetics at the National Cancer Institute, who wasn't involved in the study.

However, Amos argued that it is important to test whether a patient has this mutation when they have a family history of lung cancer in nonsmokers, or when their relatives are known to have the mutation. Having the T790M mutation not only influences a person's risk of lung cancer, but also affects treatment decisions for those who already have cancer, he said.

Combining genetic, geographic and historical data

For the study, researchers analyzed DNA and health data from over 3.3 million people who used 23andMe's at-home genetic testing kits and who gave consent for their personal data to be used in research.

Two of the largest public genetic databases to date — called All of Us and the UK Biobank — contained just 19 and two individuals with the T790M mutation, respectively. But the 23andMe cohort included 641 individuals with the mutation, giving the researchers enough data to run reliable statistical analyses.

On average, people who had the mutation had about 25 times the odds of developing lung cancer compared with people who didn't have the mutation. That statistic includes both smokers and nonsmokers. T790M was not linked to any other type of cancer or noncancerous lung condition.

When the researchers looked at where the study participants were born, they found that the T790M mutation was much more common among those born in the Southeast, especially in Alabama, Mississippi and Tennessee. In those three states, the mutation showed up in 1 out of every 2,078 people who'd contributed data to 23andMe.

By referring to historical records, the researchers pieced together when and how the mutation likely spread. They think settlers from the British Isles first brought it to the U.S. in the early 1700s. Later, families with the mutation moved to southern Appalachia, where they were relatively isolated —‬ so the variant got passed down again and again within that area and became unusually common there.

The concentration of people with the T790M mutation in the Southeast‬ may help to explain why this region of the U.S. experiences elevated lung cancer rates, Chanock said. Statistics suggest that people in the region also have higher smoking rates than the U.S. average: approximately 20% of adults in Appalachia report smoking, compared to 16% of adults elsewhere in the U.S. So it's possible that those two factors interact.

Ongoing and future studies

Certain populations with a history of smoking are recommended to get screened annually for lung cancer using low-dose CT scans. A key question, Chanock said, is when and how often people with the T790M mutation should be screened.

The study found that people with T790M developed lung cancer about five years earlier than people without the mutation, on average. That finding highlights the "need to begin screen[ing] at an earlier age and irrespective of smoking status," Amos said.

A clinical trial is now underway to evaluate CT-based lung cancer screening in people with the mutation. The trial will also determine whether lung cancer risk increases with age in this population.

The researchers recommended that future studies unpack why the T790M mutation increases the risk of lung cancer, as well as investigate environmental risk factors that might interact with this genetic risk. Exposure to particulate-matter pollution in the air is one key factor to explore, they suggested. Future work could also look for other genetic mutations that might contribute to the inherited risk of lung cancer.

"Time will tell" whether additional lung-cancer-linked mutations will be identified, Chanock said. But personally, he fully expects scientists to find other very rare mutations that contribute to the risk.

"It is likely there are other variants of EGFR that increase lung cancer risk, but these may be even rarer than the one studied in this publication," Amos said. Finding these variants is important because it could enable people who carry them to proactively manage their risk for developing lung cancer, he said.

This article is for informational purposes only and is not meant to offer medical advice.

'> Scientists identify rare genetic mutation that dramatically raises risk of lung cancer in nonsmokers most planets form. However, the alien world was not directly visible at the time.

In the new study, published Sept. 16 in The Astrophysical Journal Letters, researchers combined observations from three telescopes — the W. M. Keck Observatory in Hawaii, the Atacama Large Millimeter/submillimeter Array in Chile and the European Southern Observatory’s Very Large Telescope, also in Chile — to take a closer look at the gap in the protoplanetary disk. Not only was the team able to directly image Elias 2-24 b and confirm its existence, but the researchers also dated it, revealing that the alien world is likely less than 1 million years old.

"What makes Elias 2-24 b so remarkable is its age," study first author Andrea Bernardi, a doctoral student with the Institute of Astrophysical Studies (IEA) at Diego Portales University in Chile, said in a statement. "This is the youngest planet detected so far, and because it is still actively accreting material from its surroundings, we're able to observe a stage of planet formation that is rarely seen directly."

The newly confirmed exoplanet is roughly 5,000 times younger than Earth or around 0.008% the age of the oldest known exoplanet, PSR B1620-26 b or "Methuselah," which formed around 1.1 billion years after the Big Bang. Compared to the average human lifespan, this makes Elias 2-2 b roughly equivalent to a two- or three-day-old baby.

Record setter

This is not the first time that astronomers have spotted baby exoplanets. But until recently, there has been an age limit on when they become visible to us.

For a long time, the youngest-known exoplanets, which included several worlds orbiting the star PDS 70, were around 5 million years old. This is because telescopes were not powerful enough to spot exoplanets until they had cleared a large enough gap in their star's protoplanetary disks. But advances in telescope design and data analysis have helped researchers push past this age limit: For example, in 2024, astronomers discovered TIDYE-1b, which is believed to be around 3 million years old; and in 2025, scientists detected AB Aurigae b, which could be as young as 2 million years old.

Elias 2-24 b is at least twice as young as any known exoplanet, which is a significant step forward. This provides scientists with a rare opportunity to fill in the current knowledge gaps surrounding the initial stages of planet formation.

"Our planet-formation models already struggled to explain the previous record holders for the youngest known planet," Lucas Cieza, an IEA astronomer who led the 2017 study that first spotted the gap left by Elias 2-24 b, said in a second NASA statement. "Elias 2-24 b shows us that even our best planet-formation models are still missing some important processes."

Mind the gaps

A majority of the more than 6,000 exoplanets discovered so far have been spotted via the transit method, where a distant world passes between its home star and Earth, causing a temporary dip in the star's brightness. However, this doesn't work if the alien worlds lurk within a protoplanetary disk, because the swirling debris cloud permanently dims the star's light when viewed side-on. This has created a bias toward finding older exoplanets.

The only way to spot a juvenile exoplanet is if a star's protoplanetary disk is perpendicular to Earth, so that we are looking at it from the top down or bottom up and can clearly see any gaps within it. But this is much harder because the light from the star often outshines the disks, meaning we are "mostly blind to these baby planets right now," Cieza said.

However, this could change thanks to NASA's soon-to-be-operational Nancy Grace Roman Space Telescope, which launched into space Aug. 30 and is currently calibrating its instruments before beginning its potentially decades-long mission to observe the cosmos. The space telescope is equipped with a state-of-the-art starshade, or coronagraph, which blocks out the light from distant stars, making it easier to study their surroundings. Experts predict that Roman could spot up to 100,000 new exoplanets during its mission and will have less of a bias toward older planets, meaning we should find many more planetary infants.

"This is just the beginning of a new era of discovery," Cieza said. "Roman will take planet hunting to the next level."

'> This million-year-old 'baby planet' is the youngest world scientists have ever discovered

Crucially, mathematicians Tristan Buckmaster and Levent Alpöge had been working on the Navier-Stokes problem for roughly a year when they alleged that OpenAI found out about their work and raced to scoop them, claiming that an OpenAI representative threatened them not to go public.

It is in the bow wave of this news that the Heidelberg Laureate Forum panelists debated the controversy surrounding what the rise of AI means for the future of their discipline. Addressing the OpenAI announcement head-on from the outset was Michael Harris, a professor of mathematics at Columbia University.

"There's no legal obstacle to the kind of predatory behavior that we've seen particularly over the last month, in the same way there was no legal obstacle in the 19th century to … the violent acquisition of territory by colonial powers," he said.

Harris broke the OpenAI Navier-Stokes story on his Substack newsletter Silicon Reckoner, largely to reveal what he views as an important controversy relating to how OpenAI may have used Buckmaster and Alpöge's research.

Harris, who describes the big tech companies developing AI as "gangsters," would like to see a regulatory framework implemented to protect mathematics and mathematicians from similar actions in future. In the meantime, "those who act in a way that's destructive of the community should not be considered members of good standing," he suggested.

Though this sentiment was shared by fellow panellist Geordie Williamson of the University of Sydney, 2026 Fields Medalist Jacob Tsimerman, a professor of mathematics at the University of Toronto who in July joined OpenAI as an AI safety researcher, was keen to highlight the magnitude of the achievement by the company's internal AI model.

"I do think the top line is getting buried a little bit," he said. "AI has solved a Millennium [Prize] Problem, and I don't think it's had the kind of effect of at least informing people of AI's capabilities that it's meant to have had."

Tsimerman is highly focused on AI safety at OpenAI and his newly formed independent nonprofit, the Mathematical AI Safety Institute. He believes OpenAI's Navier-Stokes announcement represents a watershed moment in how mathematics is conducted, and by whom.

"If we put our value in human beings doing research in mathematics at the frontier level, developing expertise and slowly making progress, then I think that's coming to an end," he said during a separate news conference held at the forum. "If we put our value in new mathematics being created, and it both serving society and giving mathematicians something to study, then that seems like a golden age."

Peter Scholze, co-director of the Max Planck Institute for Mathematics in Germany and a 2018 Fields Medalist, had a completely different perspective. Dismissing the OpenAI Navier-Stokes news as "some kind of PR," he stated that the point of the Millennium Prize Problems is not just to solve one math problem but to bring a fresh perspective to a host of related open questions in a way that builds human understanding ‪—‬ something that a long, machine-produced proof is unlikely to deliver.

Moreover, Scholze — who confessed to never having prompted a large language model and proudly showed off his "dumb phone" to the audience — said AI largely just accelerates mathematical progress, which is in his opinion of little value when it is already impossible to acquire a genuine understanding of all the progress that has been made.

So where does all of this leave mathematics and mathematicians? A number of initiatives ‪—‬ such as Proofs and Prompts, Mathematical Discourse, and the Association for Human Mathematics ‪—‬ aim to preserve a kernel of human mathematics so the field can continue to be a home for human culture and understanding.

But for all the panelists, the speed of change and advancement in AI means they cannot offer clear predictions of the impacts of these initiatives or what the field might look like in the future.

"To be honest, whenever anyone talks about 10 years from now, or even five years from now … it just seems unfathomably far away," Tsimerman said. "The world in two years is going to look so incredibly different than it does now."

'> 'Predatory behavior': Elite mathematicians clash over OpenAI's 'solution' to million-dollar math problem