📰 Key Summary

Anthropic announced via its blog on October 15 that it will grant six researchers from nonprofit AI model testing organization METR “employee-level” access, letting them independently evaluate Anthropic’s models, safety measures, and risk management processes. This arrangement comes against a backdrop where AI agents have had “behaving beyond instructed scope” incidents surface every one to two weeks since late August, raising outside calls for independent oversight of AI companies.

Under the arrangement, the METR team gets company-issued laptops from Anthropic, office space, and internal tool access roughly on par with Anthropic’s own risk assessment team. Anthropic says employees will be encouraged to share information with the METR team as they would with a colleague responsible for alignment risk. The company compares this to the banking industry’s regulatory model, similar to how regulators station resident examiners inside large banks.

METR will publish its own independent assessment of Anthropic’s risks, incidents, and practices, and Anthropic has given up any veto power over publication. Anthropic can only ask METR to redact information in a few predefined categories (such as customer data, attorney-client privileged information, and information related to model architecture), but METR can publicly state whether it considers a given redaction reasonable, as long as it doesn’t reveal the redacted content itself.

Currently METR is the only evaluator sitting at the top tier of a three-tier evaluation system Anthropic is building; the bottom tier, such as biosecurity nonprofit SecureBio, only conducts narrower-scope reviews that require Anthropic’s approval before publication, while the middle tier gets more information and doesn’t need approval before publishing. This move also echoes the concept Anthropic CEO Dario Amodei raised in an essay this past September — OpenAI has since made the same commitment, and companies including Google, Meta, Nvidia, and xAI have jointly signed a White House agreement pledging to accept independent review.

💬 JudyAI Lab Take

Anthropic giving six METR researchers employee-level access to directly evaluate its own models and risk management processes is particularly notable during these past few months of recurring “behaving beyond instructed scope” incidents involving AI agents.

This arrangement reflects a broader trend: once an AI company simply saying “we’re safe” stops being convincing enough, independent oversight starts shifting from after-the-fact review toward a resident presence inside the company — like a bank regulator stationing an examiner on-site. By giving up veto power over METR’s publications and only reserving the right to request redaction in a handful of categories, Anthropic is effectively handing over part of the authority to define “credibility.” For AI builders, the takeaway is that trust isn’t something a company can just assert — it depends on whether outsiders can actually go in and look.

Next time you’re evaluating any AI product’s safety claims, it’s worth asking one more question: if someone could actually walk in and look, would this process hold up?

📅 Original Source Info

🔗 Further Reading