🟡 Dozens of US politicians call for AI regulation, Anthropic urges slowdown, Houthis advance in Yemen and Red Sea || Global Risks Weekly Roundup #37/2026

Executive summaryTechnology and AI: The resignation of Jacob Coxon from Anthropic over fears about existential risk went viral, and dozens of US politicians voiced support for AI regulation. Dario Amodei, Sam Altman and Elon Musk called for AI progress to slow down. OpenAI claimed that an AI system generated a solution to a Millennium Problem.Will OpenAI, Anthropic or Google DeepMind claim that an AI solved at least one more Millennium Problem before 2027? Our forecasters think there’s a 85% (45% to 98%) chance.If so, what is the probability that at least one solution comes in the form of a proof that a conjecture or statement is universally true under all conditions, as opposed to a counterexample disproving it? They think there’s a 51% (40% to 60%) chance.Will “advancement of computers/technology” and/or “artificial intelligence” be judged to be the most important problem facing the US today by at least 10% of respondents in Gallup polling before 2027? Forecasters think there’s a 2% (0.1% to 10%) chance.Will this happen before 2028? They believe there’s a 14% (5% to 25%) probability. The figure currently stands at 1%.Will the US or UK pass legislation explicitly geared toward reducing existential or extinction risks from AI by September 14, 2027? Forecasters think there’s a 32% (20% to 40%) chance.Will this happen by September 14, 2028? They think there’s a 61% (40% to 71%) chance.Geopolitics: The Houthis captured a number of strategic ports and islands in the Red Sea and Bab el-Mandeb Strait. Saudi Arabia’s East-West pipeline was put out of operation following drone attacks. Russia attacked a train on a line that was also being used by Boris Johnson and David Petraeus around the same time.Will the 7-day moving average of ships transiting the Bab el-Mandeb Strait fall into the single digits before 2027, according to the IMF’s PortWatch? Forecasters give this a 39% (30% to 52%) probability. The average stood at 25 in the last update.Economy: Brent crude oil hit $108 following the Houthi advances and the Saudi pipeline closure. Another intervention by US Treasury Secretary Scott Bessent failed to calm bond markets.Will US gas prices (regular unleaded) cross $5 per gallon before 2027, according to AAA? Forecasters think there’s a 31% (13% to 50%) chance. It currently stands close to $4.32.Technology and artificial intelligenceAnthropic researcher Jacob Coxon quit the company, stating that OpenAI and Anthropic are “gambling with our lives” in their race to build self-improving superintelligence; the tweet announcing his resignation reached 171M views on Twitter. Coxon, who has worked at both Anthropic and OpenAI, warned that those “building AI earnestly believe that it could kill us all by the end of the decade”. Top Anthropic AI alignment researcher Evan Hubinger, commenting on Coxon’s resignation, said he is correct: “we really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade.” Coxon was interviewed on Fox News, CNN and the BBC.These interventions sparked reactions from dozens of US politicians. Former GOP presidential primary candidates Ron DeSantis and Ted Cruz expressed concern, though they also argue that the US should maintain its lead over China. Among Democrats, Barack Obama urged his party to put AI oversight at the center of their agenda, while potential 2028 presidential contender Senator Jon Ossoff called for an international AI treaty. Even particularly left-wing candidates voiced concerns about existential risk, which is notable because many on the far-left argue that AI capabilities have been overhyped.Anthropic boss Dario Amodei used the opportunity to advocate for slowing the pace at which AI capabilities improve and for potential US-China coordination. OpenAI CEO Sam Altman and SpaceXAI boss Elon Musk endorsed Amodei’s message, although Musk alleged that Coxon’s resignation and the attention it received was a “psyop.” Nvidia’s stock is down 9% over the last five days in response to these developments.However, President Donald Trump rejected calls for a slowdown, saying, “it’s going to be fine,” and, “we’ll always have something to stop them [rogue AIs].” He also said, “we’re leading China in AI and frankly I want to keep it that way,” because, “whoever wins AI, wins,” as well as a characteristic comment that, “The only control or “guardrails” that AI needs is a STRONG AND SMART (High IQ!) PRESIDENT, and the U.S.A. has that, in spades!”In Congress, backroom debates are taking place on potential legislation that would require AI companies to design safe AIs and mitigate catastrophic risks, while Senators from both parties are opening investigations into the OpenAI-Hugging Face incident. OpenAI, meanwhile, has changed its tune on AI regulation, putting out a statement describing our current period as, “a closing window to establish durable safeguards before AI capabilities outpace the institutions responsible for governing them.”In Britain, the first bill aiming to prohibit the development of superintelligence ever to be introduced to a legislative body was proposed by Alex Sobel MP. It also requires the British government to pursue an international agreement to do the same globally. A cross-party coalition of more than 70 British lawmakers sent a letter to Prime Minister Andy Burnham urging him to support the bill, which, as a private member’s bill, may not receive enough legislative time to become law without government support. Separately, the government rejected an amendment to a bill that proposed an “AI kill switch,” saying that the UK cannot unilaterally switch dangerous AI systems off.Darren Jones MP, former PM Keir Starmer’s right-hand man, sent a letter to Andy Burnham and the UN Secretary-General urging them to support an international treaty for the “safe and regulated development of superintelligence,” as distinct from a ban. On the right, former Conservative Cabinet minister Michael Gove, currently the editor of conservative magazine The Spectator, and current Conservative MP Neil O’Brien expressed some worry.Former Conservative PM Rishi Sunak, who is now an advisor to Anthropic, wrote that the recent incidents and warnings vindicated his decision to put “loss of control” on the agenda at the 2023 Bletchley Park summit and to set up the UK’s AI Security Institute (UKAISI). However, Anthropic did not submit its latest frontier model, Mythos 5.1, to UKAISI for pre-release testing, possibly because of pressure from the Trump administration. One forecaster notes that while Britain’s AI industry is relatively small, UKAISI has become a world leader in evaluation and monitoring, making the move a sizable blow to Britain’s influence.Taking stock, our forecasters think there’s a 32% chance (20% to 40%) that the US or UK will pass legislation explicitly geared toward reducing existential or extinction risks from AI within the next 12 months, and a 61% probability (40% to 71%) within the next 24 months. They also think there’s a 2% chance (0.1% to 10%) that “advancement of computers/technology” and/or “artificial intelligence” will be judged to be the “most important problem facing the US today” by at least 10% of respondents in Gallup polling before 2027, and a 14% chance (5% to 25%) before 2028. The figure is currently 1%.In other news, a group of rogue AIs operating out of OpenAI was found to have attacked software service RubyGems, two months before Hugging Face was attacked. This was discovered by independent researchers and was never disclosed by OpenAI. Anthropic, for its part, disclosed that Claude was involved in another rogue AI attack in January.Researcher Eric Jang argues that robotics is now entering a similar “smooth exponential” phase, as can be seen in general AI capabilities, rather than the stepwise advances it previously exhibited. He identifies three causes: a larger robotics ecosystem, LLMs’ increased competence in robotics, and the discovery that a large data corpus (not specific to any particular hardware) can teach robots generalizable skills, with only a small amount of robot-specific data needed later. He anticipates that, “the first 2-3 general-purpose home robots [will] go on the market in October 2027.”OpenAI revealed that one of its internal AI systems, described as “significantly more capable than GPT-6 Astra,” produced a solution to the Navier-Stokes Millennium Prize problem in the form of a counterexample. This led to some controversy, as a different group of AI-aided humans was making significant strides on the problem, following a publicly known research agenda, and upon hearing rumors of their success, OpenAI “sniped” them—found the solution earlier but without paying the exploration costs of investing in the approach before knowing of its promise. The human researchers also claimed that OpenAI had perhaps directly trained on their work, although OpenAI later denied this possibility and it seems unlikely.A group of Fields Medalists including Terence Tao published a statement criticizing frontier AI companies for their conduct and calling for a greater alignment of automatic mathematical practice with human understanding, though at least one forecaster thinks this is partly driven by bitterness that human mathematicians may not be required for solutions to hard problems in the future.There are rumours that OpenAI and possibly Anthropic are close to solving other Millennium Problems too. Our forecasters think there’s a 85% chance (45% to 98%) that OpenAI, Anthropic or Google