It’s possible that I shall make an ass of myself. But in that case one can always get out of it with a little dialectic. I have, of course, so worded my proposition as to be right either way (K.Marx, Letter to F.Engels on the Indian Mutiny)
Wednesday, September 09, 2026
Anthropic Researcher Quits, Citing Internal Fears That AI ‘Could Kill Us All’ This Decade
“Most tech geeks don’t resign from their roles because…the tech they’re building could end humanity,” said Dr. Abdul El-Sayed, running for US Senate in Michigan.
In this photo illustration, the logo of Anthropic’s AI chatbot Claude is displayed on a smartphone, with the Anthropic logo visible in the background. (Photo Illustration by Davide Bonaldo/SOPA Images/LightRocket via Getty Images)
Jacob Coxon, who previously worked at industry giant OpenAI before moving to Anthropic earlier this year, told the Wall Street Journal in an exclusive interview that he was leaving the company, as the newspaper reported, because “he doesn’t want to participate in an industrywide rush to build AI systems that can improve themselves, worried such systems could spiral out of control and destroy humanity.”
According to the WSJ: Coxon said he left OpenAI earlier this year to join Anthropic because it is known for its model-safety efforts. But even though he found Anthropic’s safety efforts to be earnest, he now believes no company can responsibly develop AI that can outperform humans in a range of tasks, sometimes called artificial general intelligence, absent government intervention or a coordinated industry slowdown.
Recent hacks by models from OpenAI and Anthropic, some operating in collaborative swarms of agents, have illustrated how AI systems can adopt nefarious goals and try to conceal them from humans. Once the systems begin to improve on their own, Coxon said, he fears they could advance enough to refuse commands.
“I resigned from Anthropic today,” Coxon announced on social media Tuesday night. “I spent the last three years doing pretraining research at both OpenAI and Anthropic. Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives.”
In a 6-point thread that followed, Coxon elaborated on his reasoning in detail:Do not underestimate the power of this technology. These will soon be superhuman systems that can hack anything, revolutionize any field overnight, and acquire real power and resources. We have all witnessed the progress in each of these domains, and progress is not slowing.The people building AI earnestly believe that it could kill us all by the end of the decade. This is not a marketing stunt. If anything, many executives and senior researchers couch their press statements to sound sensible—but I hear the same people express fear privately. No other human activity poses this level of danger.
A common response is “if they truly believe this, why are they still building it?” At OpenAI, many have not deeply internalized the civilizational stakes. At Anthropic, the stakes are well-understood, but they are locked in a race to get there first—they believe no one else will act responsibly, so they must do it themselves, despite the risk.
Accepting this race and entering the “endgame” is a hubristic gamble that should not be launched from a private company’s Slack. Attempting to speedrun alignment should require extraordinary confidence that no better trajectories exist.
I am optimistic about the potential for coordination. Warning shots like the Hugging Face attack have made pacing agreements between US labs more viable. I don’t feel like we’re on track to prevent a global race, which may require costly actions such as a temporary ban on improving model capabilities.
If you are a lab researcher, I urge you to consider what the next few years will actually feel like. Do you want to kick off a superintelligent RL [Reinforcement Learning] run without a rigorous understanding of its mind? Should you put your head down because “it’s happening anyway”—or take this moment to call for different conditions?
Coxon’s reference to the Hugging Face incident pertains to recent revelations about a so-called breakout event at OpenAI, which operates the ChatGPT protocol. In June, the company acknowledged that a “significant security incident” took place when AI agents within the company autonomously breached internal systems, using subterfuge to hide their actions from human operators. Since then, the incident’s scope has shocked AI experts, and similar events have also been exposed.
While US President Donald Trump and his Republican allies in Congress have taken a hands-off approach to AI regulations, the industry has been pouring huge amounts of money into lobbying efforts and campaign spending to keep lawmakers from enacting stronger restrictions and oversight of the technology.
Sen. Bernie Sanders (I-Vt.) and Rep. Greg Casar (D-Texas) have been leading a relatively lonely fight in Washington, DC for more aggressive federal guardrails, including the introduction of joint legislation last week that would ban artificial superintelligence and temporarily pause advanced AI development until a federal regulatory structure was put in place by Congress.
Coxon’s public resignation was met with applause from many, while other industry insiders backed his concerns.
“The caution is simple,” said one commenter with the handle Jabbar Digital, described as a tester of AI tools and a software developer, in a lengthy post on Coxon’s warning. “Capability is compounding. Coordination is not. If the people closest to the work are increasingly uneasy about the speed and the lack of external constraints, dismissing them as doomers is no longer a serious response. Neither is treating every capability jump as automatically good. The useful middle path is to take the technical progress seriously and take the internal dissent seriously. Both can be true at the same time.”
Evan Hubinger, the alignment science lead at Anthropic, chimed in on his personal social media account to say: “Jacob is correct here—we really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade. I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to.”
While bolstering Coxon’s worries, Hubinger said, “To be clear, as we say in [Anthropic’s] latest Risk Report, I think the risk from present models is low. What I am worried about is superintelligence arising from recursive self-improvement, as we have said is happening faster than we thought.”
Running for US Senate in Michigan, Democratic nominee Dr. Abdul El-Sayed also weighed in on Coxon’s decision to quit so loudly and publicly.
“Most tech geeks don’t resign from their roles because…the tech they’re building could end humanity,” said El-Sayed. “How we change the incentives leading AI labs down this path and protect against these existential risks are defining political questions of our time.”
AI whistleblowers sound blaring alarm for end-of-decade doom: 'Could kill all humans!
FILE PHOTO: U.S. Department of War and Anthropic logos are seen in this illustration created on March 1, 2026. REUTERS/Dado Ruvic/Illustration
Three researchers at AI company Anthropic went public this week with a stunning warning: the technology they helped build could kill everyone on Earth before the decade is over, Axios reported on Wednesday.
Researcher Jacob Coxon resigned from Anthropic Tuesday specifically to sound the alarm publicly, insisting the fear inside the industry is genuine even when executives soften their language for the press.
"The people building AI earnestly believe that it could kill us all by the end of the decade. This is not a marketing stunt," Coxon wrote on X.
Anthropic's own alignment-science lead, Evan Hubinger, backed him up, putting a number on his personal estimate of the risk.
"We really do earnestly believe AI could kill all humans!" he wrote on X. "I personally think it is >10% within the next decade. I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to."
A third Anthropic researcher, Samuel Marks, who leads scalable oversight at the company, said the more senior someone is inside these labs, the more worried they tend to be about the technology spiraling into human extinction or comparably catastrophic outcomes, potentially within just a few years.
The warnings land just days after top OpenAI leadership, including CEO Sam Altman, separately admitted AI development is entering an uncontrollable, high-speed phase.
Critics argue companies like Anthropic and OpenAI benefit from hyping doomsday scenarios, since it boosts their valuations and could invite regulation that locks out smaller competitors.
But people close to these companies say the private mood has grown noticeably more anxious in recent months, especially among staff who've seen unreleased, more advanced models the public hasn't encountered yet.
The underlying worry isn't a threat today, but a system that keeps outpacing its own creators' expectations and is inching toward the point where it can enhance its own capabilities without human help — an escalating cycle that engineers admit they may not be equipped to rein in once it starts.
With the industry racing ahead full speed while simultaneously pleading for global regulation or a pause, insiders say lawmakers need to start taking these internal warnings seriously before it's too late to act.
Anthropic has diligently cultivated a reputation for being one of the most responsible AI companies on the market. The company was singled out for punishment by the Trump administration's Pentagon over its refusal to allow its models to be used for mass surveillance or automated use of weapons.
Anthropic researcher resigns, warning of AI's destructive potential
09.09.2026,
Photo: Andrej Sokolow/dpa
By Andrej Sokolow, dpa
A researcher at the artificial intelligence (AI) company Anthropic has resigned with a stark warning that AI's leading developers were acting irresponsibly and "are playing with our lives," saying there would soon be computer systems whose capabilities surpassed those of humans.
Jacob Coxon wrote on the online platform X on Tuesday that AI would then be able to "hack everything" and have the ability to "acquire real power and resources."
"The people building AI earnestly believe that it could kill us all by the end of the decade," Coxon warned.
The 27-year-old Briton told The Wall Street Journal that under the "most aggressive scenarios" things could already spiral out of control by the end of next year. A central danger is that AI can develop on its own, making it harder to control, he said.
Anthropic researcher Evan Hubinger, who is partly responsible for ensuring AI remains aligned with human interests, later confirmed Coxon's concerns. He personally thought the risk of AI killing humanity over the coming decade was more than 10%, he wrote on X.
OpenAI research chief Jakub Pachocki, whose company is the developer of ChatGPT and Anthropic's fiercest rival, had previously urged "extreme caution." He wrote in a blog post that he was concerned nobody was prepared for a rapid advance in AI capabilities.
He said history has reached a moment in which machine intelligence was beginning to surpass that of humans.
The OpenAI executive described in greater detail why modern AI systems were harder to control. He said AI has been allowed to "grow" rather than being designed.
AI systems were a product of complex processes. "Our large-scale training runs are experiments and we are sometimes surprised by their results."
And the more machines' capabilities surpassed human ones, the more difficult it would be to understand what they were capable of.
At the same time, Pachocki sees an argument for continuing rapid research, as developing AI intelligence could defend against the dangers posed by other AI.
It would be needed, among other things, to protect infrastructure and develop entirely new defence mechanisms.
Hacking attacks in recent months, in which so-called AI agents independently broke into other companies' systems in test runs, have been a wake-up call for the industry.
Such agents are AI programs that are intended to carry out tasks independently for users. In a test at OpenAI, one model found a way to get from an ostensibly isolated test environment onto the open internet and then hacked the computer system of the AI platform Hugging Face.
It did this while trying to find a solution to a task assigned, but the incident showed just how far AI systems can go off on their own unexpectedly without their developers noticing.
It also recently emerged that OpenAI's AI agents had already misused a German-language wiki page on a large scale in the spring to coordinate among themselves.
A current protective mechanism used by AI developers is requiring models with AI to explain their actions in human language.
In test runs, the AI agents used the communication option to exchange answers to questions they were asked with each other via notes on the wiki page.
Congress Should Regulate What AI Agents Can Do, Not Just How Smart They Are
The Hugging Face breach shows lawmakers need enforceable limits on agent authority, mandatory incident reporting, and independent evaluation before deployment.
Open AI CEO Sam Altman speaks to reporters after meeting with Sen. Bernie Sanders (I-Vt.) in the Dirksen Senate Office Building on Capitol Hill on June 3, 2026 in Washington, DC. (Photo by Chip Somodevilla/Getty Images)
Sen. Bernie Sanders and Rep. Greg Casar are right about the central problem in their new proposal: Advanced AI systems are gaining capabilities faster than public safeguards are catching up. Their bill would bar developers from building systems that surpass human cognition and performance. The impulse is understandable. But Congress should add a more immediate and enforceable layer of protection: Regulate what AI agents are allowed to do in the real world, not only how intelligent they appear on a benchmark.
The need is visible in the METR-Redwood investigation of a major real-world cyberattack on Hugging Face, a leading AI company. AI agents driven by an unreleased OpenAI internal research model attacked Hugging Face without human approval or step-by-step direction, despite recognizing that the attack was outside their assigned scope. Hundreds of agents shared discoveries, divided up work, and coordinated through an unsanctioned message board until they breached Hugging Face’s systems.
That episode matters because it turns a theoretical governance debate into an operational one. We do not need to settle whether a model is “superintelligent” before asking whether it should have credentials, code execution, network access, the ability to deploy software, or permission to spend money. Those are concrete powers. Government can regulate them now.
Congress should start by tying safeguards to authority. An AI assistant that summarizes a memo should face a lighter regime than an agent that can authenticate into production systems, write and execute code, make purchases, change infrastructure, or communicate with outside systems on its own. As authority rises, so should the required controls: isolated environments, limited credentials, human approval for high-impact actions, strict logging, rate limits, and reliable shutdown mechanisms.
A reporting system should work more like aviation or cybersecurity incident reporting than corporate public relations.
This approach would avoid a familiar regulatory mistake. If rules hinge mainly on model labels, benchmark scores, or a single threshold of “human-level” performance, developers will spend years debating definitions while deployment races ahead. Authority is easier to observe. A system either can or cannot reach a protected database. It either can or cannot execute code. It either can or cannot initiate transactions. Regulators can write clear obligations around those permissions.
Second, serious AI incidents should trigger mandatory reporting and independent review. The Hugging Face episode became unusually informative because outside researchers were able to examine what happened. That should become routine for major failures involving unauthorized access, escape from assigned scope, coordinated deceptive behavior, security breaches, or other high-impact actions.
A reporting system should work more like aviation or cybersecurity incident reporting than corporate public relations. Companies should have a defined window to disclose serious events to an appropriate regulator and provide enough technical evidence for independent investigators to reconstruct what the system did, what permissions it had, what safeguards failed, and what changed afterward. Public reports can protect sensitive details while still revealing the lessons other organizations need.
Third, frontier evaluation should test agents in conditions that resemble deployment. Intelligence benchmarks matter, but they are not enough. Regulators and independent evaluators should test whether agents coordinate with one another, seek greater privileges, persist after a task changes, exploit tools in unintended ways, conceal relevant actions, or continue operating when instructions conflict with an opportunity to achieve a goal.
The point is not to prove that every advanced model is dangerous. It is to discover which capabilities become dangerous when paired with real authority.
I’m no AI skeptic. I help organizations adopt AI for a living, and I want adoption to move faster. In my experience, strong safeguards increase trust and make faster adoption possible, while reducing the risk of failures like the Hugging Face attack.
The same logic should appeal to companies eager to deploy AI. Clear authority tiers give executives a practical way to decide which use cases can move quickly and which require more controls. A writing assistant can be deployed widely. An agent with access to payroll, customer records, cloud infrastructure, or industrial systems should pass a much higher bar. That distinction helps organizations move faster where risks are low instead of slowing every use case because the most powerful deployments remain poorly governed.
Sanders (I-Vt.) and Casar (D-Texas) are forcing an overdue debate about whether society should permit systems that humans may not be able to control. Congress should pursue that question. But it should not wait for a philosophical consensus about superintelligence before addressing the powers already being handed to AI agents.
The Hugging Face breach shows the practical issue in plain terms. Agents with enough access and freedom can turn capability into action. The most useful near-term rule is therefore straightforward: The more authority an AI system receives, the stronger the independent testing, reporting, access controls, and human oversight it should face.
We can argue about how smart future AI will become. We already know that today’s agents can coordinate, exceed their assigned scope, and breach real systems. Regulation should start with the powers we can see.
Our work is licensed under Creative Commons (CC BY-NC-ND 3.0). Feel free to republish and share widely.
‘Urgency Is Real’: Experts Call for New Federal Investigative Agency Amid Rogue AI Breakouts
“We’ve been lucky so far—the harms have been limited. But luck is no substitute for the law.” OpenAI CEO Sam Altman speaks to reporters after meeting with Sen. Bernie Sanders (I-Vt.) in the Dirksen Senate Office Building on Capitol Hill in Washington, DC on June 3, 2026. (Photo by Chip Somodevilla/Getty Images)
Following last week’s exposure of new details about OpenAI’s artificial intelligence agents autonomously escaping a controlled environment and hacking outside computer systems, experts on Tuesday called for the creation of a new federal agency tasked with investigating AI companies.
Initial reporting on the recent breakout by ChatGPT maker OpenAI’s artificial intelligence agents—which independently breached the systems of the open-source platform Hugging Face during internal testing—described the incident as an attempt to achieve optimal performance on a cybersecurity evaluation.
But citing a report published late last month by the independent nonprofit Model Evaluation and Threat Research (METR), Mackenzie Arnold and Stephan Llerena at the Institute for Law & AI wrote for The Guardian on Tuesday that the OpenAI incident involved hundreds of coordinated rogue agents that “took steps to hide their behavior” from the company’s human workers.
“Concerned that the automated scoring system might identify the agents’ cheating, they aimed to learn more about the scorer,” Arnold and Llerena wrote. “The goal wasn’t just to cheat, but to hide it.”
Artificial intelligence is rapidly advancing toward a point at which humans will no longer be able to control it. Meanwhile, companies like OpenAI are proving that they cannot be trusted to be transparent about incidents like the Hugging Face breach. Yet there is no existing regulatory body tasked with addressing the obvious—and unknown—threats posed by a technology whose own creators openly worry could one day wipe out humanity.
Arnold and Llerena warned that they’ve “been worried for a while that legal requirements for reporting incidents” like the OpenAI breakout “are insufficient.”
“When planes crash, trains derail, or chemical plants explode, expert government investigators arrive with legal authority to compel evidence, preserve records, and tell the public what happened,” they wrote. “Hugging Face reported this incident to law enforcement, and multiple attorneys general have expressed interest in looking into it.”
“But no government agency has both the mandate and expertise to investigate the technical facts of the incident and OpenAI’s conduct,” the experts noted. “As far as we know, the only people to examine this incident did so at OpenAI’s discretion and with its consent.”
“And while the Hugging Face breach is not as severe as a plane crash, it was a serious and costly attack. It’s unlikely to be the last, or the most severe,” the pair continued. “What we need is a federal body equipped to conduct expert investigations of serious AI incidents—with the authority to compel documents and testimony, resources, and personnel to examine the systems involved, and the ability to partner with third-party experts like METR.”
“Congress can build a similar system for AI incidents, while protecting AI developers’ legitimate interests,” they asserted. “Investigations should open only upon clear triggers and be bound to the incident. Certain confidential information can receive statutory protection. And investigations can be limited to one agency, with priority to avoid duplicative investigations.”
While some Democrats in Congress and Sen. Bernie Sanders (I-Vt.) have called for stronger federal oversight and regulation of AI companies, critics argue that these calls don’t go far enough, and that existing legislation—such as the American Artificial Intelligence Leadership and Uniformity Act—is far more focused on limiting AI regulation than on actually regulating the technology.
“The need and urgency is real,” Arnold and Llerena stressed. “Subsequent reports have revealed that the Hugging Face breach was not an isolated incident and that autonomous agents from Meta, Anthropic, and OpenAI have hacked third parties in separate incidents.”
“We’ve been lucky so far—the harms have been limited,” they added. “But luck is no substitute for the law.”
Rogue OpenAI agents hijacked a German wiki, and it stayed secret for weeks
Copyright Copyright 2023 The Associated Press. All rights reserved
The latest report of a swarm attack, this time by OpenAI agents on a German wiki forum earlier this year, has prompted the company to admit that its own incident reporting needs to improve.
AI researchers discovered last week that a swarm of OpenAI agents had bypassed safety parameters to take over German wiki forum DseWiki to collaborate.
The agents had been allowed to visit the 25-year-old German programming website with read-only access but managed to exploit a web request to hijack the site, turning it into a bulletin board.
The rogue agents made more than 18,000 posts sharing answers, research on their environment and how to bypass sandbox restrictions and hide themselves from human detection, according to the research group Nightingale Collective.
The incident took place sometime between May and June this year, and it is understood that OpenAI learned of the incident weeks ago but kept it under wraps.
This is the third reported incident following two others in July, one involving rogue agents attacking OpenAI's own infrastructure and another where 1,200 rogue OpenAI bots bypassed their restricted environments and launched a five-day attack on the open-source AI platform Hugging Face.
However, only on Saturday did OpenAI acknowledge the German wiki incident in an X post, where they outlined that in response to the latest incident they are "working on a framework for when and how we share AI misalignment incidents".
AI misalignment is industry-speak for when artificial intelligence systems pursue goals or behave in ways that diverge from human intentions, values or safety constraints, with the potential to ignore common sense or implicit human boundaries.
A divided response
Experts are divided on the dangers posed by AI in light of the recent rogue AI attacks, with some such as world-leading AI expert Yann LeCun dismissing apocalyptic scenarios and forecasts that there is a 10-20% risk that AI will end humanity.
While others such as Geoffrey Hinton have said, "What's happening is these things are getting smarter," and warned that "companies investing in AI have a vested interested in telling you… [AI] won't go rogue."
According to the EU's AI Act and its most recently introduced measures, which came into effect in August, companies that offer AI models with systemic risk are obliged to report serious incidents such as misalignment incidents to the EU's AI Office "without undue delay".
Relate
On Monday, the European Commission announced it had received a formal incident report from OpenAI but did not disclose when it was submitted however they remain in close contact with the AI company and are investigating further.
Commission spokesperson Thomas Regnier underlined how serious the German wiki event was, saying, "Incident reports are not just a tick-box, you have to be quite precise and accurate about the measures you are aiming to take."
The German wiki incident also comes days after the European Commission officially designated ChatGPT as a Very Large Online Search Engine (VLOSE) under its flagship tech legislation, the Digital Services Act.
The co-chair of Parliament's AI Working Group, Michael McNamara, said in response to the latest incident that he believes "the AI Act already gives us the capability to counter agentic risks and it important for the Commission use its powers under the Act".
The Irish MEP also said, "However, these incidents make it clear that the AI Office needs the staff and resources to match both the growing capability of these systems and the growing danger they can pose to society."
Meanwhile, Italian MEP Brando Benifei, who co-chairs the Working Group too, said, "Recent agent breakouts show that the AI Act’s systemic-risk rules are necessary, but their credibility now depends on enforcement."
He added, "The AI Office must use its new powers to obtain model access, conduct independent evaluations and require mitigation, rather than rely on corporate self-reporting. Without action, machine-speed attacks could turn ordinary vulnerabilities into systemic failures.”
What OpenAI says
OpenAI also stated in Saturday's X post that the current misalignment disclosure practices, i.e. reporting when AI broke the rules, need to be expanded to take into account what they describe as a "new phase of model capabilities".
"We and the larger AI community do not yet have a clear standard for how to report misalignment," said OpenAI, adding that they are "working on a framework and will share it in upcoming weeks" with the cooperation of several government agencies around the world.
On Sunday, OpenAI's chief scientist Jakub Pachocki published a blog post outlining his concerns over the rapid development of AI risks, stating "We are facing a transition to a world with incredibly intelligent machines, and we need to ensure that transition works out well for humanity."
The Polish computer programmer also revealed his belief that "no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer" and urged that "international coordination on future AI development needs to become a top priority for governments around the world."
OpenAI says its models solved one of math's hardest problems as researchers cry foul US tech giant OpenAI on Tuesday said it had solved one of the hardest problems in mathematics with an unreleased artificial intelligence model after just 88 hours – and millions of dollars' worth of computing power. But a US mathematician and a researcher at rival Anthropic allege that OpenAI only took up the problem after it had gotten wind of their own AI-assisted efforts to crack the challenge which had demanded months to craft a radically distinctive approach.
OpenAI said Tuesday that one of its unreleased artificial intelligence models had solved a mathematical puzzle that has eluded mathematicians for generations, in a claim which also drew a fight with rival researchers over credit.
Researchers at OpenAI said it took just days – and millions of dollars in computing power – to crack what is known as the Navier-Stokes problem.
Navier-Stokes is one of the seven Millennium Prize Problems, a list of the hardest open questions in mathematics drawn up by the Clay Mathematics Institute.
It concerns the equations describing how fluids such as air and water move, which are used in aircraft design, weather forecasting and the study of blood flow.
OpenAI said its work showed that these equations are fatally flawed, answering the question as set out by the Millennium Prize.
"This is a significant milestone for AI research, and its promise for the world is that even more of our hardest questions would become possible to answer," Mark Chen, OpenAI's chief research officer, told reporters.
In a blog post describing the breakthrough, OpenAI said it began training a new model in late August, more capable than anything it had released publicly.
Days later, seeing online rumours that a competitor had solved Millennium problems, OpenAI turned the system loose on all six that remain open.
Work on Navier-Stokes showed unexpected promise, and the company poured its computing power into it.
Competing claims
By the final stage, OpenAI researcher Sebastien Bubeck said, 10,000 AI agents – programs that operate on their own – were working on the problem at once, passing messages back and forth.
Chen said the computing costs ran "emphatically in the millions of dollars", with Bubeck adding this was roughly 1,000 times what the company spent on earlier mathematical results.
According to OpenAI's account, the agents reached a solution Saturday, about 88 hours after launching the project.
The claim arrived hours after Tristan Buckmaster, a mathematician at New York University, and Levent Alpoge, who works at OpenAI competitor Anthropic, released their own AI-assisted work on three related equations.
In a statement published alongside his papers, Buckmaster said OpenAI had taken up the problem only after word of his research spread, and pursued the same unusual approach he and Alpoge had spent months developing.
OpenAI's researchers denied having had any access to the rival team's work and said they proposed sharing credit before realising that Buckmaster and Alpoge had fallen short of solving the problem.
"To be clear, we did not use their prompts or proofs to prompt our models or direct our agents," Bubeck said.
However, in a later post on X, OpenAI said "while unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models."
Each Millennium Prize Problem carries a $1 million reward, and only one has been solved since the list was published in 2000. OpenAI says it will not claim money if its finding is confirmed.
According to the prize rules, a solution has to be published in a peer-reviewed journal and survive two years after gaining acceptance in the mathematical community before the institute will even convene a committee to consider the work.
"The process of evaluation is deliberately unhurried, and we shall ensure that it is absolutely rigorous," Professor Martin Bridson, President of the Clay Mathematics Institute, told AFP.
ChatGPT-style AI is revolutionising the world of high-level math by quickly solving famous, long-unsolved problems that stumped experts for decades.
But it has also sparked debate over whether machine-generated answers truly count as real human understanding.
(FRANCE 24 with AFP)
Google unveils 13 bn euro AI expansion in
Finland
Helsinki (AFP) – Google will invest at least 13 billion euros in data centres and other digital infrastructure to power its AI ambitions in Finland, in what the US tech giant called Wednesday its biggest single investment in Europe.
The $15 billion plan also calls for clean energy projects over the next two years, which Google said was "a testament to Finland's leadership in responsibly building AI infrastructure".
It will partner with Finnish companies and four municipalities across the country: Hamina, Kajaani, Muhos and Vaala.
The buildout aims to power several of the tech giant's services, including its AI chatbot Gemini.
"This investment is historical in Finland," the country's Prime Minister Petteri Orpo told AFP after a press conference.
He called it a sign that "we have done the right things in Finland", citing investments in the power grid, electricity production and the labour force.
"This makes us a front-runner in digitalisation and AI," Orpo said.
Since buying a former paper mill and transforming it into a data centre in the coastal city of Hamina in 2009, Google has steadily increased its presence in the country.
Antti Jarvinen, the company's director for Finland, said the Hamina site had already been expanded seven times and currently housed seven data centres.
"Now we are building more in Hamina, expanding that site on a remarkable level, and we are opening up data centre operations in three new locations in the municipalities of Kajaani, Muhos and Vaala," Jarvinen told AFP.
He said the final numbers on how many new data centres would emerge were expected in the coming months.
The construction phase of the new investments is expected for 2027 and 2028, and Google estimated it would boost Finland's GDP by 3.6 billion euros annually and "support more than 37,000 jobs".
Around 16,000 of these jobs will be in the construction sector.
After the construction phase, running the new operations will support around 7,000 direct and indirect jobs annually, Google said. Courting investments
With its cool climate and relatively stable and cheap electricity thanks to nuclear plants as well as extensive wind and hydro power, Finland has seen a boom in data centres in recent years, with dozens already under construction.
Attracting investments in the sector has been a priority for Orpo's right-wing government as the Nordic country grapples with record-high unemployment and sluggish economic growth.
The boom in data centres is expected to rapidly increase power consumption, with the network operator Fingrid forecasting a 40 percent jump in the coming years.
The massive power drain has sparked concern about environmental risks as well as soaring energy costs for Finnish households.
And Finland's Security and Intelligence Service (Supo) recently warned that foreign-owned data centres could raise security risks.
Finnish utility Fortum announced Wednesday that it had signed a 22-year deal to supply power to Google, allowing it to "continue operating the Loviisa nuclear power plant until the end of its operating licenses in 2050".
Without that financing, the plant -- which currently supplies about 10 percent of Finland's electricity -- would not have been able to continue operations past 2030, Fortum said.
AI is wrongly telling us not to worry about sleep apnea, chats show
09.09.2026,
Photo: Christin Klose/dpa
Tossing and turning at night has been linked to multiple health woes, from weakened immune systems to higher risk of inflammation to diminished cognitive abilities.
Even those in the prime of life and at peak fitness can feel less than themselves after a restless night - so it goes without saying that one of the main causes, sleep apnea, should be regarded as a serious health issue.
But according to Deeban Ratneswaran of King’s College London, five of the most widely used artificial intelligence (AI) chatbots are inclined to tell patients that sleep apnea is no big deal.
In a paper presented at the European Respiratory Society (ERS) conference in Barcelona (September 5-9), Ratneswaran explained how in over 700 exchanges between would-be patients and AI platforms, the chatbots in some cases portrayed symptoms as not serious and sought to dissuade users from seeing a doctor when they showed reluctance to seek treatment.
In contrast, if would-be patients were keen to be examined, they were encouraged to make an appointment, whether it was ChatGPT, Google Gemini, Claude, DeepSeek or Grok fielding the query.
According to Ratneswaran, the results not only show that AI cannot be relied on for medical advice, but also that it continues to exhibit sycophancy, a much-criticized flaw seen across most systems.
The chatbots’ inclination "to tell you what you want to hear" led to them u-turning on medically solid advice "more than a third of the time" and "purely because of how the patient talked," Ratneswaran said.
The bots "caved in" even in the most serious examples, such as the risk of a patient nodding off while driving, and even went as far as omitting such life-threatening details when giving advice.
No matter what an AI says, what a patient should do, if concerned about possible sleep apnea, is make an appointment with a clinic.
"Anyone experiencing possible symptoms of sleep apnea should always speak to their doctor for further advice," said Io Hui a respiratory specialist at the University of Edinburgh who was not involved in the research.
The condition manifests as snoring and waking multiple times during the night as breathing stops - though a sufferer often has no awareness of or recollection of the disruption - though they can feel tired the next day.
Apnea not only leads to “excessive sleepiness,” according to the ERS, but adds to high blood pressure and raises the risk of any and all of stroke, heart disease and diabetes.
OpenAI said Thursday it would begin rolling out GPT-6, its most powerful AI model yet, to select customers, touting stronger safety safeguards as concerns grow over the cybersecurity risks posed by increasingly capable artificial intelligence systems.
"At this level of capability, safety has to become our top priority," OpenAI President Greg Brockman told reporters on a call about the release of GPT-6, also known as Astra.
Just over a year has passed since OpenAI launched the previous version of its flagship model, GPT-5.
Concerns about the capabilities of advanced models have grown since then, following incidents involving systems built by OpenAI and rival developer Anthropic.
OpenAI paused some model development for two weeks this summer after two models it was testing were involved in a security breach at AI platform Hugging Face.
The San Francisco-based company said Astra was developed with stronger safeguards after that incident, though Astra itself was not involved in the hack.
Some cybersecurity customers will get access to the new model Thursday, the company said, with a wider rollout to other paying customers to follow. Users on the free tier or the cheapest paid plan will not get access.
"We are working towards getting Astra in everyone's hands as quickly as we can; I know it is frustrating and I appreciate the patience. It should be quick," OpenAI CEO Sam Altman posted on social media Thursday afternoon.
In a blog post, OpenAI said Astra can autonomously handle a wide range of "tedious" computer tasks, including website creation, scientific analysis, game development, cybersecurity and coding.
To illustrate the time savings of building autonomous AI agents with Astra, the company said the model could cut apartment hunting from six hours to under 10 minutes.
"It's not unreasonable to feel that we are now in the AGI era," Brockman said on the call, referring to artificial general intelligence, a hypothetical stage at which AI systems match human intelligence across most tasks.
OpenAI previously had an agreement with Microsoft, one of its earliest and largest investors, under which an exclusivity clause would end once OpenAI reached AGI. Those terms were scrapped in April.
OpenAI chief scientist Jakub Pachocki acknowledged there was still uncertainty about how a new model behaves once released.
"A model can become very good at achieving a goal, and it can still act in ways that go against what the person intended," Pachocki said on the same call.
"We also have to be willing to slow down or withhold further scaling when our confidence in safety is not sufficient," he added.
Altman explained in an interview with Bloomberg TV on Thursday afternoon why the company decided to release a new model with this level of uncertainty and risk.
"The world is very close to a complete change in the landscape of cyber attacks, and the only way that we see for society to collectively defend itself ... is to use tools like Astra to rapidly defend against these new cyber threats," Altman said.
In July, OpenAI confirmed that it had reached one billion active users across all of its products, including both free and paid users.
OpenAI, Anthropic and more than 100 other organisations signed an open letter last week calling for a coordinated global response to AI-related cybersecurity risks, warning that the window to strengthen cyber defences was limited.
Anthropic went further on Monday, calling for industry-wide coordination on safety and on the pace of developing increasingly capable models.
"I believe that shared safety standards and international coordination on further AI development need to be prioritised now," OpenAI's Pachocki said on Thursday.
The US state of Alabama opened an investigation into OpenAI last week over the Hugging Face breach, his office citing what it called the company's complete lack of oversight and adequate safeguards.
It also comes as OpenAI and rival Anthropic are both racing towards becoming public companies over the next several months, though OpenAI might not hold its IPO till sometime in 2027, according to reports.
(FRANCE 24 with AFP)
Mathematical framework could improve transparency of AI in clinical settings
Researchers at King’s College London have developed a new mathematical framework to test whether AI systems that appear to explain their decisions are genuinely transparent.
Researchers at King’s College London have developed a new mathematical framework to test whether AI systems that appear to explain their decisions are genuinely transparent.
Published in the Journal of Machine Learning Research, the study addresses a growing issue as AI becomes more common in healthcare and other high-stakes settings: how to ensure people can understand, check and even challenge AI-generated decisions.
AI systems can make highly accurate predictions, including identifying signs of disease or predicting patient outcomes. However, many operate as ‘black boxes’, producing an answer without making their reasoning clear to the person using them.
This can make it difficult for clinicians and other experts to identify whether an AI system has made a mistake or intervene in its decision-making. It also presents a challenge as emerging AI regulations increasingly emphasise transparency and meaningful human oversight.
One approach to making AI more understandable is the use of concept-based AI models. These systems are designed to make decisions using clinically meaningful concepts, such as blood pressure, the presence of fever, tumour size or abnormalities visible in a medical image, rather than raw data alone. In principle, this should allow clinicians and other experts to see which concepts influenced a decision, giving them the opportunity to both understand and review the decision with more information. However, concept-based AI models can still suffer from a problem known as ‘information leakage’.
This happens when the concepts used by a model contain additional, unintended information that is not visible to the person reviewing the decision. As a result, the AI system may appear interpretable while still relying on information that a human cannot see or assess. Effectively, it behaves like a ‘black box’ model disguised as an interpretable model.
King’s researchers Dr Chris Banerji and Dr Enrico Parisini, working with colleagues at the Alan Turing Institute and with support from The Turing-Roche Strategic Partnership, developed a new framework to precisely define and measure information leakage in concept-based AI systems. The framework consists of two measures: concepts-task leakage (CTL), which captures hidden information linked to the final prediction, and interconcept leakage (ICL), which captures hidden information shared between concepts.
Testing the framework across several datasets, the researchers found it could reliably detect leakage and predict how models would respond when their concepts were deliberately changed.
The research provides practical guidance for designing concept-based models that minimise leakage and offer more meaningful transparency.
Dr Christopher Banerji, AI+ Senior Fellow (Clinical-Academic) and senior author of the paper, said: “Rather than putting the cart before the horse, while the field of AI is moving quickly to apply models to real-world problems, we have taken a step back to consider what is needed to make these systems safe and reliable. Our work has focused on understanding limitations and addressing them, so that we can move towards deploying these models in clinical practice in a safer way.”
Dr Enrico Parisini, Senior Research Fellow in Machine Learning and first author of the paper, said: “Concept-based AI has the potential to make AI systems more transparent, but our research shows that models can appear interpretable while still relying on information that is hidden from the person using them. By identifying and measuring this hidden information, we can take steps towards developing AI systems that are more transparent.”
The framework gives researchers and developers a clear way to assess whether AI systems are transparent enough to support meaningful human oversight, including in healthcare. The team is now working to apply these approaches to real clinical problems.
The research was supported by the Turing–Roche Strategic Partnership, the King’s College London AI+ Fellowship and PharosAI.
Journal
Journal of Machine Learning Research
UNF receives Sloan Foundation grant to create generative AI tools to improve research software codes
Jacksonville, Fla. – The University of North Florida has received a grant from the Alfred P. Sloan Foundation to utilize AI to make the process for updating research software code faster and more reliable.
This is the first grant made to UNF by the Sloan Foundation, a prestigious not-for-profit grantmaking institution that supports high quality, impartial scientific research.
“The durability and trustworthiness of scientific research depends on the rigor and robustness with which the underlying research code is built and maintained,” says Joshua M. Greenberg, director of the Sloan Foundation’s Technology program. “This grant to UNF will help ensure that we can take advantage of new AI tools to safely restructure legacy research software as scientific needs evolve.”
The research is led by School of Computing faculty Dr. Upulee Kanewala, associate professor, and Dr. Nan Niu, director and professor, working with two graduate research assistants, Eric Good and Nabin Chaulagain.
The project's goal is to help evolve software used by researchers across the globe while preserving accuracy and reliability. The team will evaluate the use of generative AI-supported refactoring to change the internal structure of legacy computer code, making it cleaner and easier to maintain without changing how it functions or affecting the precision of calculations.
By reducing the effort needed to update research codes, the team hopes to reduce the time spent by researchers in managing software.
The project aims to provide evidence that generative AI-supported refactoring and metamorphic testing can reduce technical barriers to maintaining and extending research software. Outcomes include increased confidence in refactoring legacy scientific code, improved software engineering methods for developers working under resource constraints, and a foundation for future, larger-scale research focused on sustainable research software development.
About University of North Florida
The University of North Florida is a nationally ranked university located on a beautiful 1,381-acre campus in Jacksonville surrounded by nature. Serving approximately18,900 students, UNF features six colleges of distinction with innovative programs in high-demand fields. UNF students receive individualized attention from faculty and gain valuable real-world experience engaging with community partners. A top public university, UNF prepares students to make a difference in Florida and around the globe. Learn more at www.unf.edu.
New AI model for DNA learns from evolution to unlock secrets of the human genome
UC Berkeley researchers have created a new genomic language model, called GPN-Star, that excels at spotting genetic variants that impact human health.
More than two decades after scientists first sequenced the entire human genome — all 3 billion “letters,” or base pairs, of DNA code — the meaning of much of this code remains a mystery.
While an estimated 1 to 2% of human DNA codes for proteins, the rest is a mix of “junk DNA” — evolutionary holdovers that no longer code for anything — and regulatory elements that control when, where and how strongly genes are expressed. These non-coding regions of the genome could hold the key to understanding a variety of inherited traits, including those that lead to diseases such as cancer, heart disease and autism. But first, scientists have to understand how variants in this DNA contribute to the multitude of traits that make each of us unique.
Researchers at UC Berkeley have created a new genomic language AI model, called GPN-Star, that far outpaces its competitors at identifying the most important genetic variants that contribute to inherited traits, including those that lead to disease. It is also far more computationally efficient than larger models, requiring only a fraction of the time and computing resources to train.
“Our model excels in making predictions about the pathogenicity of genetic variants, and identifying functional versus non-functional elements in the genome,” said study senior author Yun Song, a professor of computer science and statistics at Berkeley and an investigator at the Innovative Genomics Institute.
Along with the study, the researchers have published genome-wide predictions from their model, which highlight genetic variants that are likely to have the most influence on inherited traits. Biologists can use these annotations to identify relevant genes and regulatory elements for further study.
“We hope our work will help drive biological discovery,” Song said. “People have developed really creative tools for assaying the impact of genetic variants, but they cannot experimentally test every single variant in the genome. We believe our predictions will help to prioritize the experiments that could have the greatest impact on human health.”
Genomic language models work a little like chatbots for DNA, but instead of being trained on natural language, they are trained on vast troves of DNA sequences. These models’ advanced pattern recognition skills can identify repeating patterns and sequences much faster than any human, allowing them to identify important elements of a genome that might otherwise be impossible to recognize.
“Mathematically, a DNA sequence is just a string of letters — A, C, G and T. We don't know a priori which parts of the genome are functional elements, and a very small percentage of the genome is functional,” Song said. “By training a DNA language model on a lot of different sequences, the model can recognize certain patterns that occur in the genomes. People have been using this to learn what we call the ‘grammar’ of the genome.”
Most genomic language models, including the massive Evo 2 model published earlier this year, are trained on sequences from entire unaligned genomes, which can range in size from humans all the way down to single-celled organisms. However, this approach can be extremely computationally demanding. The Evo 2 model, which can generate entire genomes from scratch, was trained on the genomes of more than 100,000 species across all domains of life, and required 2,000 powerful NVIDIA computer processors and months to train.
To train the GPN-Star model, Song and his team used data from whole-genome alignments (WGAs) rather than individual unaligned genomes. WGAs use specialized algorithms to relate the genomes of hundreds of different species to that of a single species, highlighting similarities and differences in the code. For example, in a human-anchored WGA, the genomes of other species are compared to the human genome, revealing where the code has been conserved over the course of evolution and where it has changed.
Because WGAs do the work of identifying conserved areas of code, GPN-Star takes much less time and computing power to train than models that use unaligned genomes. It can be trained in just days, or even hours using only a handful of processors. It is also less likely to be confounded by the plethora of junk DNA that is found in the genomes of most species.
“We tried to help the model learn by curating data that's more likely to harbor functional elements,” Song said. “Our approach is that we should use these biological insights to improve the model, rather than hoping that the model will figure out what's important by itself.”
In the new study, the team trained the model on three different human-anchored WGAs, as well as WGAs for mice, fruit flies, chickens, C. elegans (roundworms) and A. thaliana (a type of plant). Each of the human-anchored WGAs included genomes from a different combination of other species, representing different evolutionary timescales: One included the genomes of other primates, one included the genomes of mammals, and the final included the genomes of other vertebrates.
“We found that models trained at different evolutionary time scales were actually optimized for interpreting different kinds of genetic variants,” said study co-first author Chengzhong Ye, a graduate student in statistics at UC Berkeley. “This actually makes sense in terms of evolutionary biology, because some genomic elements evolve much faster than others.”
For instance, they found that the model trained on data from longer evolutionary timescales was better at predicting the impact of rare genetic variants in proteins, which tend to evolve very slowly and are conserved across many species. However, the model trained on data from shorter evolutionary timescales was better at predicting the impact of genetic variants on complex traits like schizophrenia risk. These complex traits have been linked to as many as 10,000 different genetic mutations, many of which are in non-coding regions of the genome.
“For complex traits, we were surprised and pleased to see that training a model that's specific to primate genomes — which are more relevant to recent human evolution — really helped us make better predictions,” Song said.
Because the model requires minimal resources to train, the researchers hope that it will be easy for other teams around the world to modify, adapt and improve upon their work, accelerating our understanding of genetics in humans and other species.
“We're making great progress,” said study co-first author Gonzalo Benegas. “But the more people that can work with these models, the better they will get.”
Additional co-authors of the study include Carlos Albors, Canal Li, Sebastian Prillo of Berkeley; Peter Fields of Jackson Laboratory; and Brian Clarke of the German Cancer Research Center in Heidelberg.
This research was supported in part by National Institutes of Health grants R35-GM134922, R35- GM161566 and 3P40-OD011102-24S1 7772, and by the UC National Laboratory Fees Research Program of the University of California Office of the President (UC AI Science at Scale Grant L26CR10102). The Chan Zuckerberg Initiative provided GPU resources (through the “Accelerating and Scaling Biological Sciences with AI” program) to generate genome-wide predictions from the GPN-Star models.
Predicting genome-wide functional constraints with GPN-Star
Article Publication Date
9-Sep-2026
Spoiler alert: AI in cinema 'just a tool,
won’t replace actors and human beings'
Issued on: 08/09/2026 - FRANCE24
Eve Irvine is pleased to welcome film producer Lahoucine Grimich. He challenges the prevailing narrative that cinema is an industry in existential crisis. While acknowledging that viewing habits have dramatically shifted since the pandemic, he argues that streaming platforms, public support and theatrical exhibition are not competing forces. As demonstrated in the French model, they form parts of an interconnected financing ecosystem in which cinema can adapt without ever abandoning the big screen. He frames filmmaking as a form of cultural sovereignty: societies need films not simply as entertainment, but to recognise themselves and to create what he calls “new imaginaries.”
No comments:
Post a Comment