OpenAI dismisses three safety researchers as outside evaluators step forward

OpenAI fired three safety researchers in a dispute over AI risks². Separately, Anthropic is removing its model evaluations from the internet, though not permanently, according to a report published on 10 October⁴.

At the same time the assessors are becoming visible. CNBC reported that AI’s quiet safety gatekeepers are stepping into the spotlight, in a piece illustrated by Anthropic chief executive Dario Amodei at the White House¹. Another account reported that independent AI evaluators are gaining importance as the debate over safety and regulation grows, and said the position was backed by AI leaders including OpenAI’s Sam Altman, in a passage the excerpt leaves unfinished³.

Why it matters to investors

The industry’s stated preference is to police itself. In Silicon Valley, AI founders cheered President Trump’s calls for the industry to police itself on safety, after months of debate about leading AI companies’ unchecked agents interfering with everything from US government systems to a list the excerpt does not complete⁵. The same report was carried with a photograph of Trump beside Nvidia chief executive Jensen Huang and Elon Musk in the East Room on 29 September⁶.

That sits oddly beside what the same executives are asking governments to do. Musk, Altman and Anthropic’s leadership have pressed world governments over the pace of AI development, calling for a slowdown⁷. So the labs are asking for external restraint in public while one of them narrows outside access to its own test results^4,7. If safety claims become a condition of selling to governments and regulated industries, who signs them off is a commercial question rather than a philosophical one^1,3.

What to watch

Whether the evaluators get standing. Being in the spotlight is not the same as being in a statute or a procurement rule, and the reporting so far records attention rather than authority^1,3.

Then the incident record. A timeline of developments in AI safety since the attack on Hugging Face has now been published⁸, which gives an outside reader a way to date the sector’s claims. Third is the personnel signal: a lab that dismisses safety researchers during a dispute about risk² and a lab that withdraws its evaluations from public view⁴ are both reducing the material an outsider can check.