Anthropic and OpenAI pledge to embed outside safety evaluators

In an essay published over the weekend, Anthropic CEO Dario Amodei proposed embedding third-party evaluators inside frontier AI companies, giving them power to report safety incidents, assess whether models are truly aligned, and publish findings without editorial control from the company. Amodei said Anthropic would commit to giving independent evaluators such as METR and Redwood Research unprecedented access to its systems, and OpenAI CEO Sam Altman said his company would commit to the practice too.

Evaluators who spoke with TechCrunch welcomed the proposal but said the details need to be worked out — ideally backed by legislation — before they can know whether they’ll function as independent watchdogs or as vendors operating on the AI companies’ terms. Historically, outside reviewers tested finished models shortly before release; evaluators now want access to intermediate training checkpoints, the post-training environment that rewards certain behaviors, and evaluation transcripts and logs, plus the ability to interview employees to check whether public safety claims match internal practice.

Neither Anthropic nor OpenAI has said which evaluators they’ll work with, when they’ll be embedded, how many will be brought on, or what can be disclosed publicly. Past access has often been limited: OpenAI gave METR and Redwood roughly a week on premises to investigate the Hugging Face incident, and Apollo Research was given only three days to test OpenAI’s GPT-6 Astra before release — in both cases, too little time to draw firm conclusions, the evaluators said.

Adam Gleave, CEO of FAR.AI, said his firm has turned down contracts with several frontier developers that wanted too much control over the evaluation process; by default, he said, evaluators are treated as ordinary contractors bound by restrictive NDAs. So far, Meta, SpaceXAI, and Google DeepMind have not committed to embedding third-party evaluators, though DeepMind CEO Demis Hassabis has proposed a separate industry standards body.

Some regulation is emerging: California’s SB 53, signed into law last year, requires large frontier developers to publish safety frameworks and report critical incidents, and SB 813, signed this month, creates a framework for state-recognized “independent verification organizations.” The EU AI Act already requires frontier developers to document evaluations and report serious incidents.

Henry Papadatos, executive director of Safer AI, said voluntary measures depend entirely on a company’s goodwill. “You cannot have it both ways, having zero accountability externally, and then say, ‘I’ll just have my own flexible rules,’” he said.

Why it matters

If Anthropic and OpenAI actually cede control over what evaluators can access and publish, it would mark a rare shift toward outside scrutiny of frontier AI development — but both companies have a track record of tightly limiting evaluator time and access, and neither has committed to specifics yet.