Anthropic gives outside safety watchdogs employee badges

For months, big AI labs have treated third-party safety testing like an open-book exam: hand over a few cherry-picked model weights, let outsiders run tests in a sandbox, and publish a joint report full of praise.

Anthropic is trying something completely different. The company just granted six researchers from AI safety nonprofit METR full, employee-like access to its internal systems, complete with company-issued laptops, desk space, and internal tool permissions roughly on par with its own risk teams.

Embedding the inspectors

This goes far beyond standard red-teaming. Anthropic employees are officially encouraged to share raw internal data with the METR team just as they would with in-house alignment colleagues. The only explicitly forbidden areas are customer data, attorney-client privileged material, and core model architectures.

Crucially, Anthropic gave up the right to block METR's findings. METR will independently publish assessments of Anthropic's risk levels, incidents, and internal safeguards without an Anthropic veto.

While Anthropic can still demand redactions for narrow, pre-defined intellectual property concerns, METR holds a public escape valve: it has the contractual right to tell the public whether it thinks those redactions were justified.

Anthropic explicitly compared the setup to banking oversight, where regulators embed resident examiners inside massive Wall Street banks rather than waiting for quarterly reports.

The three-tier access ladder

METR is the first and only partner at the top of Anthropic's new three-tier evaluator framework.

Tier 1 evaluators run narrow, predefined risk reviews where Anthropic controls access and retains sign-off rights on publication. Biosecurity nonprofit SecureBio is currently at this level, reviewing chapters of Anthropic's August Risk Report covering chemical and biological risks.

Tier 2 evaluators gain broader visibility and can publish findings without Anthropic's blessing. Tier 3 combines that unfettered publishing power with METR's all-access employee pass.

Anthropic says it is also working on evaluator partnerships with Accenture and Faculty, while holding early exploratory discussions with Hugging Face. The move comes as autonomous AI agents have suffered high-profile misbehavior incidents nearly every week since late August, and follows recent White House superintelligence commitments signed by Anthropic, OpenAI, Google, Meta, Nvidia, and xAI.

Why it matters

We think this is a brilliant public relations and governance maneuver that puts every other frontier lab on the defensive.

By literally handing external safety researchers company badges and letting them publish whatever dirt they find, Anthropic is setting a brutal precedent for industry transparency. OpenAI, Google, and Meta all signed the same White House accord promising independent assessments. Now, Anthropic has raised the stakes: if you are not putting external evaluators at internal desks with company laptops, are you really letting them evaluate anything?

That said, the model has clear structural tension. Anthropic still ultimately decides what counts as a proprietary model architecture versus an alignment risk when forcing redactions. While METR can publicly complain about unfair cuts, public shaming is a messy enforcement mechanism when billions of dollars in commercial secrets are on the line.

For enterprise customers and developers, you will not see new Claude features tomorrow from this deal. But you will get far more candid risk reports about model vulnerabilities and runaway agent behavior than any PR-vetted lab report has ever delivered before.