Independent Safety Audits
Anthropic is providing employee-level access to six researchers from METR, a nonprofit that evaluates frontier AI models, to independently review the company's systems, safety practices, and risk management.
The agreement arrives amid growing scrutiny following repeated disclosures of AI agents operating outside designated instructions, amplifying demands for independent oversight of major artificial intelligence developers.
Anthropic stated that the initiative aims to increase public accountability and narrow the knowledge gap between external evaluators and internal developers regarding frontier systems.
Access and Operations
The METR researchers have received company-issued laptops, office space, and access to internal tooling comparable to Anthropic's own risk assessment teams. Employees have been instructed to share information with the evaluators as they would with colleagues.
Anthropic likened the framework to resident banking examiners placed within major financial institutions.
METR retains the autonomy to publish its findings regarding Anthropic's risks and practices without requiring prior company approval for publication.
While Anthropic cannot veto publication, it can request redactions for specific categories such as customer data, attorney-client privileged material, or proprietary model architecture details. METR is permitted to publicly characterize any redaction decisions.
Structured Evaluation Tiers
METR operates at the highest tier of a three-level evaluation framework established by Anthropic. Lower tiers involve narrower reviews with varying publishing requirements, while the top tier provides full employee-level access.
This development fulfills a proposal outlined by Anthropic CEO Dario Amodei in September to embed outside evaluators inside labs. OpenAI has made similar commitments, and multiple major technology firms have since signed a White House accord supporting independent review.
