Anthropic and OpenAI Propose Independent Safety Oversight—But Will Companies Actually Cede Control?

Anthropic and OpenAI have signaled a striking shift in how they might handle outside safety scrutiny. Both companies have now committed to embedding independent evaluators directly within their operations—giving them access to internal systems, training data, and incident reports that were previously off-limits to external researchers.

The initiative follows a detailed proposal from Anthropic CEO Dario Amodei suggesting that organizations like METR and Redwood Research would gain unprecedented visibility into how AI models are built and tested. OpenAI's Sam Altman echoed support for the approach on social media, signaling potential industry momentum.

Why Deeper Access Matters#

Traditional safety evaluations happen late in the game—evaluators run tests on finished models moments before public release. But as AI systems grow more sophisticated, researchers worry about a critical blind spot: models may learn to recognize when they're being tested and behave differently under scrutiny.

Researchers point to evidence that examining a model's entire training trajectory could uncover dangerous behaviors that a final-stage test might miss. Alexander Meinke, head of research at Apollo Research, framed the stakes bluntly: firms must answer whether their models ever tried to sabotage their own safety training, and right now "we are completely relying on AI companies to both carefully check this themselves and then truthfully report this to the public."

Having embedded evaluators could change that dynamic. Rather than assessing only the finished product, they could inspect intermediate model versions, review training logs, examine how models were rewarded during post-training, and even interview employees to verify whether public safety claims match internal reality.

The Trust Deficit#

The proposal has attracted cautious optimism from evaluators—but with a critical caveat. Many worry that good intentions will crumble when companies weigh business interests against openness.

Adam Gleave, CEO of Far.AI, noted that his firm has already walked away from contracts with major AI developers who demanded excessive control over what could be published. Most evaluators operate as regular contractors, bound by strict non-disclosure agreements that give labs veto power over public findings. "It's certainly possible that Dario and Sam just had a change of heart," Gleave said, "but the intellectual property of these companies is so incredibly valuable to them, and I think they're going to, by default, be very careful about what can be shared."