Anthropic Lets Outside Watchdog Move In — Literally
Anthropic has given a nonprofit safety evaluator what it describes as 'employee-like access' to its internal systems, hardware, and office space. Six staff from METR, a research group that has previously worked with OpenAI, Google, and Meta, are now embedded inside the Claude developer's offices with access broadly comparable to what Anthropic's own risk teams have.
The arrangement is the first concrete step toward a commitment Anthropic chief Dario Amodei floated last September, when he managed the unlikely feat of getting Sam Altman and Elon Musk to publicly agree with him. The proposal: frontier AI labs should let external auditors move in and actually look around. OpenAI backed the idea enthusiastically at the time. It has yet to do anything about it.
Anthropics's stated rationale is straightforward enough. Its models are getting more powerful, the risk assessments are getting harder, and it wants external accountability without putting potentially dangerous technical details into the public domain. The bank supervision analogy is doing a lot of work here — the idea being that just as regulators embed examiners inside systemically important banks, third-party evaluators should be able to assess AI risk from the inside before something goes wrong.
METR is not being paid for this. The organisation refuses donations from AI companies and their employees, funding itself through individual donors, philanthropic bodies, and institutions like the UK's AI Security Institute. That funding independence is presumably the point — it does at least put some distance between evaluator and evaluated.
Still, the arrangement is not without critics. Sceptics have drawn comparisons to the accounting sector, where cosy relationships between auditors and clients have historically produced optimistic assessments at inconvenient moments. Others argue the whole model is a way for labs to manage the optics of safety while lobbying against harder government regulation.
Anthropics has tried to address the independence question structurally. METR can publish its findings without Anthropic's approval. Anthropic can require redactions in specific categories — customer data, legal privilege, architectural details that could be misused — but it cannot simply kill an unflattering report.
METR is one of the organisations that dug into the Hugging Face incident involving rogue OpenAI models, which gives it at least some credibility as a genuine investigator rather than a rubber stamp.
Separately, Anthropic is also working with Accenture on safety testing — a partnership announced last month with both sides committing at least a billion dollars — and with SecureBio, which handles chemical and biological weapons risk assessment.
The timing is not incidental. Anthropic is heading toward an IPO expected this year at a valuation north of two trillion dollars, and apparently spent a significant chunk of recent investor presentations talking about safety. Reassuring the market that existential risk concerns won't derail the listing is clearly part of the agenda.
None of this is a substitute for regulation. But as voluntary measures go, letting an independent team move into your building and read your internal risk documentation is at least more substantive than publishing a responsible AI policy on your website.