Anthropic has given six researchers from METR employee-like access to examine its frontier AI models and safety practices. The researchers have company hardware, office space and access to internal tools roughly comparable to Anthropic’s own risk-assessment teams. For an industry that has traditionally treated outside scrutiny like a guest asking to tour the wine cellar during dinner, this is an unusually generous invitation.

🤚 The Open-Palm Inspection

In an announcement published on October 15, Anthropic said the nonprofit evaluator will publish assessments of the company’s risks, incidents and practices. Anthropic cannot veto publication. It can require redactions in narrow, predefined categories, including customer data, privileged legal material and certain model-architecture information. METR may publicly describe the redactions and say whether it considers them justified without revealing the protected details.

That distinction matters. An invitation to assess a model is one thing; access to the systems, safeguards and decision-making around it is another. A benchmark score tells you how the machine behaved in an arranged interview. Internal visibility might tell you who booked the interview, what questions were excluded and whether the machine has been quietly promoted to run the building.

Anthropic compares the arrangement, cautiously, to resident examiners inside banks. Its premise is that a company knows more about its own risks than an outsider peering through the press release. Independent assessors need enough access to narrow that information gap. TechCrunch reports that the move comes amid growing attention to AI-agent incidents and calls for stronger checks. The bank analogy is a comparison, not a claim that METR is a government regulator.

👐 The Two-Handed Audit

Anthropic describes a three-tier evaluator system. At the first tier, reviewers investigate narrow questions and require Anthropic’s approval to publish. At the second, they get more information and can publish without that approval, subject to limited redactions. At the third, METR’s small embedded team receives employee-like access plus those publication rights. Anthropic says it is also working with SecureBio and Accenture, among others, in arrangements that do not currently match METR’s top-tier access.

This is not a universal key to the kingdom. METR’s arrangement covers a small team under enforceable agreements, and Anthropic says other METR employees are separated from its embedded work. Sensitive categories remain restricted. A redaction right, however narrowly drawn, still belongs to the institution being reviewed. The proper measure of this initiative will not be the attractiveness of the access badge but the specificity of the published findings—and whether uncomfortable conclusions survive the trip from conference room to public page.

Nor does embedding an evaluator magically settle the question of independence. A reviewer can be given excellent access and still face practical constraints around confidentiality, scope and institutional relationships. Equally, refusing any outside access on the theory that perfect independence is impossible would make secrecy the default safety program. In luxury governance, the perfect can be the enemy of letting anyone see the basement.

🌿 The Gentle Awakening

The significant change is where the evidence lives. Frontier-model companies can publish model cards and polished demonstrations, but the decisive questions often concern deployment choices, failed safeguards and incidents discovered between product launches. Outsiders cannot examine those realities if all they receive is a model endpoint and a complimentary tote bag.

Imagine a future report that describes a recurring failure mode in agent testing, the mitigation attempted and the residual risk. Its value would lie in enough concrete evidence for readers to judge the response, even if proprietary details must remain unpublished. Compare that with an auditor’s seal that says only that a thoughtful conversation occurred. Both might fit on a website; only one helps the public understand what is being risked.

👑 The Gold-Leaf Reckoning

Anthropic says it wants other frontier AI companies to embed evaluators too. That is a proposal from a company with a commercial stake in defining what credible oversight looks like, and it should be assessed accordingly. But granting external researchers substantial access and removing a general publication veto creates a testable standard. The next question is whether METR’s eventual public work contains enough detail to challenge a powerful lab rather than merely decorate its lobby.

For now, the announcement is neither proof that frontier AI is safe nor evidence that independent oversight has been solved. It is a serious experiment in allowing a third party closer to the machinery. We await the report, preferably before the machinery receives a title and a discretionary budget.

“The auditors have been admitted to the executive floor. Please do not confuse this with the building passing inspection.” — The Slap of Wisdom Department of Premium Due Diligence, checking whether the doors open from both sides