Six METR researchers are now inside Anthropic. They can publish whatever they find.

Anthropic onboarded six researchers from METR on October 15 with what the company is calling employee-level access: company hardware, office space, internal tools and systems, and access roughly on par with Anthropic's own risk assessment teams. The announcement introduces a three-tier framework for how external evaluators get access to frontier AI labs. Tier 3 — the level METR now holds — gives evaluators the right to publish their findings without Anthropic's approval.

That last sentence is the news. Not the access. The publication rights.

Source spread

• Anthropic — An update on independent evaluators [hype] — Anthropic's own framing. More specific than most safety announcements, but the bank-examiner analogy does a fair amount of flattering work on the company's behalf.
• METR — Model Evaluation & Threat Research [safety] — The evaluator's stated mission: understand dangerous capabilities in frontier models before deployment. METR has reviewed Anthropic's Risk Reports multiple times over the past year.
• SecureBio [safety] — Currently engaged at Tier 2 reviewing the chemical and biological weapons chapter of Anthropic's August 2026 Risk Report.

Pros & cons

What's real here:

• Publication rights without a veto is genuinely unusual. Most safety partnerships involve the company controlling what gets published. Under this framework, if METR finds something alarming and publishes it, Anthropic can require redactions in predefined categories but cannot kill the finding. And METR can comment publicly on whether those redactions were justified without revealing the redacted content. That's structural accountability. You can't un-publish it.
• The three-tier ladder creates a path for more evaluators to follow. SecureBio is at Tier 2 for the biosecurity chapter. Accenture with its Faculty division is in early stages. Hugging Face is in early discussions. Tier 1 gives scoped access with Anthropic approval required for publication; Tier 2 allows publication without approval but with redaction rights; Tier 3 is the full employee-level access METR now holds.
• This happened partly because Anthropic signed the White House Accord on Super Intelligence last month — a voluntary commitment from the signing labs to work with independent external evaluators. Anthropic went further than "work with": six people with desks and hardware, siloed from the rest of METR to prevent competitive leaks.

What deserves a side-eye:

• Six people. Anthropic has hundreds of researchers. A team of six, however capable, evaluating the safety posture of one of the most capable AI labs on the planet — the ratio is not comforting. The bank-examiner analogy holds except for the part where the Fed doesn't limit itself to six examiners per bank.
• The redaction discretion matters a lot. Anthropic can require redactions for customer data, attorney-client privileged information, and model architecture details. "Model architecture details" is the interesting one. Architecture information carries high leak risk, and determining what counts as architecture versus capability is a judgment call Anthropic is making, not METR.
• The rest of the industry hasn't moved. No Tier 3 evaluator announced at OpenAI, Google DeepMind, or xAI as of this writing. The framework Anthropic is publishing is better than the norm because the norm is weak.

What builders need to know

• Follow METR's publications. If they find something about a capability you're relying on, that's the signal that matters. metr.org publishes public reports. Worth a bookmark.
• SecureBio's bioweapons-chapter review is the near-term signal for life sciences builders. If you're building anything in healthcare or biology on Claude, watch for when SecureBio publishes its Tier 2 review of the August Risk Report.
• Nothing changes in the API today. This is an oversight framework, not a product update. What changes is who's watching, what they're allowed to say, and what happens when they find something.
• The White House Accord sets the floor. Other frontier labs signed the same accord and are going to be measured against this framework. If your stack depends on OpenAI or DeepMind, watch whether they match this level of access or stay at Tier 1 equivalents with approval-required publication.
• Redaction comments are the canary. If METR ever publishes a note saying "Anthropic required redactions that we believe were not justified," that's the moment to pay close attention. That mechanism is now real.

Further reading

• Anthropic — An update on independent evaluators — primary source, full framework details
• METR — Model Evaluation & Threat Research — the evaluator organization; prior published reports
• SecureBio — the biosecurity evaluator engaged at Tier 2