Anthropic said on October 15, 2026 that it has begun giving employee-like access to researchers from METR, the nonprofit known for probing frontier systems. The group plans to publish its risk findings. The move follows Anthropic's recent pledge of stronger third-party assessment and its signature of the White House Accord on Super Intelligence, which called for work with outside auditors.
Six METR staff now have company laptops, desks and internal tools that Anthropic says roughly match what its own risk-assessment teams use. Outside a short list of excluded categories such as customer data, attorney-client privileged material and certain model-architecture details, staff are told to brief the visitors the way they would brief a colleague on alignment risk.
Anthropic argues that powerful models need harder safety evaluation, but that not every relevant fact can be published without helping attackers or exposing commercial secrets. The company's answer is to give trusted evaluators deep access and let them publish reasoned conclusions. It lists four requirements for such partners: frontier evaluation expertise, independence from equity or operational dependence on Anthropic, enough access to see training and safeguard details, and freedom to publish.
Publication rules matter. Anthropic says it cannot veto METR's analyses of risk levels, incidents or practices, though it can require redaction of narrowly defined sensitive material. TechCrunch's same-day report emphasises the same employee-level framing and notes the timing amid a stretch of disclosed agent mishaps that have sharpened calls for outside checks.
METR is not Anthropic's only evaluator track. The company says it is also working with SecureBio and Accenture, and is in early talks with other groups including Hugging Face. Over the past year METR had already reviewed several of Anthropic's Risk Reports; the new arrangement deepens that relationship from document review toward resident access.
Anthropic compares the setup, carefully, to bank supervision, where resident examiners sit inside large institutions because outsiders cannot see the full risk picture from public filings alone. It presents the AI version as an early step toward stronger safety oversight rather than a finished regulatory regime.
Anthropic is explicit that the residency is part of a learning process across multiple evaluator partnerships rather than a one-off publicity move. How quickly METR's first public write-ups appear, and how much technical detail survives redaction, will decide whether outside researchers treat the programme as genuine transparency or as controlled access theatre.
What the first METR publications will cover, how redaction disputes will be handled in practice, and whether other frontier labs will match the employee-like access standard remain open. For now the verified core is narrower and still consequential: a named independent evaluator has been given deep, ongoing access inside Anthropic with a commitment to publish.