Anthropic CEO Dario Amodei has committed the company to embedding third-party safety evaluators with employee-like access, making outside oversight the first concrete step in his proposal to slow advances in frontier AI capabilities. TechCrunch reported the plan on September 12.

In his essay, “We Must Pace the Frontier,” Amodei says evaluators should verify safety commitments, report incidents and assess training pipelines as well as completed models. He names METR as an example of an evaluator organization, not as a confirmed appointment.

Access beyond a finished model

Anthropic intends to invite an external review team with office desks, badges and company laptops. Reviewers would receive access to workspaces and tools mostly comparable to internal risk-assessment teams, with exceptions for legal obligations, contracts and private customer or partner information.

The proposed contract would let reviewers publish key findings about risks, incidents, practices and access without Anthropic's editorial control. The company would retain narrowly defined redaction rights for sensitive information, but not simply because a finding is unfavorable. Reviewers could publicly flag redactions that affect their conclusions.

A commitment, not an industry agreement

Amodei's wider framework calls for coordination among frontier developers in democratic countries, followed by efforts toward international agreements. He says some industry discussions would need government support because of antitrust constraints.

Those broader steps remain proposals. The essay describes the external team's arrival as planned for the near future, rather than documenting a completed deployment. Amodei also explicitly distinguishes pacing from halting model training: the stated aim is to give alignment work, safeguards and independent verification more time to keep up with capability gains.