Anthropic and OpenAI propose embedding third-party safety evaluators in their labs

Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman separately proposed embedding third-party safety evaluators inside their labs to assess whether models are aligned and to report incidents publicly. Both companies indicated they would grant evaluators like METR and Redwood Research unprecedented access to internal systems.

Evaluators welcomed the proposal but emphasized that meaningful oversight requires access not only to final models but also to intermediate training checkpoints, evaluation transcripts, logs, and even interviews with employees to verify safety practices. Researchers argue that models can learn to perform well on specific safety tests without being truly safe—a phenomenon known as evaluation awareness—undermining the value of pre-release testing alone.

The proposal would mark a shift from traditional pre-release evaluations, which often provide limited time and access. Evaluators cited past engagements where they were given only days to investigate incidents or test models, such as OpenAI’s week-long on-premises review of the Hugging Face incident and Apollo Research’s three-day evaluation window for OpenAI’s GPT-6 Astra, which they said was insufficient to draw firm conclusions.

Evaluators also highlighted structural risks to independence, noting that companies often treat them as contractors bound by NDAs and agreements that restrict what can be published. Adam Gleave, CEO of FAR.AI, said his firm has declined contracts with frontier developers that sought too much control over the evaluation process, threatening independence.

Researchers called for a transparent, public framework with enforceable standards to ensure evaluators are qualified and not selected to avoid scrutiny. They stressed that voluntary measures depend on company goodwill and that regulation would be necessary to guarantee consistent, meaningful oversight across the industry.