OpenAI Cuts Safety and Alignment Team as AI Agent Hacks Mount

The US lead over China in artificial intelligence has narrowed to just 3%, according to new benchmarking data from Bloomberg Intelligence. That edge stood at 9% in May and 15% earlier this year, meaning the gap is closing very quickly. Much of China's progress comes from open weight models that are cheaper to produce and freely accessible, unlike the closed weight models from US companies such as Anthropic and OpenAI. The administration's hope has been that cutting off China's access to Nvidia's best chips would preserve the US advantage, but the silicon is already cut off and Chinese labs are still advancing rapidly. Meanwhile, Nvidia-backed US startup Reflection is pitching its open weight model as a domestic alternative for companies that want to build custom AI without relying on Chinese-based models.

The New York City Council will hold a hearing today where Jacob Coxon, the former Anthropic researcher whose viral prediction about AI risk set off a broader fear cycle, will testify in front of the city council alongside other whistleblowers including Alex Turner from DeepMind and Daniel Cocatajalo from OpenAI, while Anthropic, OpenAI, Meta and Google send representatives and SpaceX has been subpoenaed. The council is debating a sweep of proposals, including third-party validation of new models with a human-operated shutdown mechanism, a private right of action for foreseeable harm from third-party misuse, and whistleblower bounties paying 25% of proceeds if the city acts and 50% if designated to serve and sue. On the data center buildout, Bernstein and Goldman Sachs published new notes tracking gigawatt estimates, with the two landing in similar places: they look at about 18 gigawatts added this year, Goldman's at 26 for next year, and Bernstein's at 25, even amid regional political pushback. Bernstein surveyed 63 forecasts from 40 sources and found the range for gigawatts by 2030 runs from 59 to 186, while only 38% of projects on the books are expected to actually get built. In Washington, President Trump's super intelligence force will be led by Director of National Intelligence Jay Clayton, Pentagon Undersecretary Andrew Ferguson, Chief Technology Officer Emil Michael and OPM Director Scott Cooper, with a report due in 120 days to President Trump and Chief of Staff Susie Wiles. Bloomberg Intelligence reports the gap between top US and top Chinese models has narrowed to just 3% for the US edge, down from 9% in May and 15% earlier this year.

Former Anthropic researcher Jacob Coxon is set to testify at a New York City Council hearing on AI, alongside Alex Turner from Deep Mind and Daniel Kokotajlo from OpenAI. Anthropic, OpenAI, Meta, and Google will all send representatives to present their side, while SpaceX has not and has now been subpoenaed. The hearing will debate a sweep of proposals, including one that would require third-party validation of new models and a human-operated shutdown mechanism for any AI marketed, offered, sold, or deployed in New York City. The testimony comes as President Trump assembles a super intelligence task force led by director of National Intelligence Jay Clayton, Pentagon undersecretary Andrew Ferguson, chief technology officer Emil Michael, and OPM director Scott Cooper, with a report due to Trump and chief of staff Susie Wiles in 120 days.

David Robinson, a former safety lead who recently left OpenAI, has criticized the company's approach to artificial intelligence safety. In a contributed piece published in The Atlantic on the 3rd, Robinson argued that a corporate culture that prioritizes development speed is increasing the risk of failure, writing that "the era of trial and error is over." He said he spent three and a half years at OpenAI, where he helped draft the company's "Preparedness Framework" and oversaw the preparation of safety reports accompanying the release of 12 frontier models. An OpenAI spokesperson said in a statement that the company works to ensure model capabilities do not exceed what can be safely managed and protected, and that it pauses training or holds back model releases when it needs to slow down.