Most AI safety claims are checked by the company making them. On October 15, 2026, Anthropic announced a step in the other direction: it is giving a six-person team from METR, a nonprofit that evaluates frontier AI models, what it calls "employee-like access." Anthropic says it cannot veto what METR publishes about it.
That is a notable promise, but the details matter. This post walks through what was announced, how the three-tier system works, what stays off limits, and the caveats METR itself has raised.
Quick summary: On October 15, 2026 Anthropic said six METR employees get company hardware, office space and internal tools roughly comparable to Anthropic's own risk-assessment teams. METR will publish its analysis, Anthropic says it cannot veto publication but can require redactions in narrow predefined categories, and customer data, privileged legal material and model architecture are excluded. All of this is vendor-reported and voluntary.
What was announced
According to Anthropic's post, six METR employees receive Anthropic hardware, office space and internal tools and systems roughly on par with those available to Anthropic's own risk-assessment teams. Anthropic staff are encouraged to share information with them just as they would with a fellow employee working on alignment risk. (Source: Anthropic)
Only that small team works on the arrangement under enforceable agreements, and other METR staff, including those on other Anthropic projects, are walled off from it, per Anthropic. The company says this is not a way for other organizations to obtain Anthropic's or its customers' information. Anthropic frames the context as a commitment made last month to stronger third-party assessment and its signing of the White House Accord on Super Intelligence.
The no-veto rule, and its limits
The headline claim is publication. Anthropic says METR will publish its analysis of Anthropic's risk levels, incidents or practices and that Anthropic cannot veto that publication. However, Anthropic can require redactions in predefined categories, such as reproducing confidential intellectual property, and Anthropic decides what falls into those categories. METR can say publicly that a redaction happened and whether it thinks it was justified, as long as it does not reveal the redacted content. (Source: Anthropic; FourWeekMBA)
Excluded from sharing are customer data, attorney-client privileged information and model architecture details. Anthropic's stated reasoning is that information about architectures "presents a high risk of inadvertent leaks, and its relevance to risk assessment is typically low," so it is generally withheld unless highly relevant and then needs explicit internal approval.
The three tiers
| Tier | Access | Publication | Example |
|---|---|---|---|
| Tier 1 | Narrow risk investigations; Anthropic chooses access | Only with Anthropic's approval | SecureBio (chem/bio chapter of August 2026 Risk Report) |
| Tier 2 | Broader information access | Without approval; Anthropic can require removal of specific categories | Evaluators who meet the criteria |
| Tier 3 | Employee-like access, six staff | Similar to Tier 2 | METR |
Anthropic also lists criteria for evaluators: expertise, independence (no equity and no financial or operational dependence on the company), sufficient access and freedom to publish. It says it is working toward a partnership with Accenture and Faculty and is in early talks with Hugging Face and others. (Source: Anthropic)
What METR itself says
The story is not purely celebratory. FourWeekMBA's summary of METR's own October 14 post says METR accepts no payment or donations from frontier AI companies, and that a voluntary regime makes it hard to ensure "adequate independence to collect and publish evidence," because METR still depends on companies to grant access and permit publication. (Source: FourWeekMBA)
Anthropic compares the setup to bank regulators' resident examiners while noting the analogy is imperfect, and says: "We're embedding evaluators voluntarily, and we believe every frontier AI company should do the same." The sources reviewed do not disclose the engagement's duration, payment terms, the date of METR's first published assessment under it, or which Anthropic models METR will assess.
Why this matters to everyday users
- It tests whether safety claims can be checked by someone who does not work for the vendor.
- It sets a template other labs could be asked to match, especially as governments discuss incident reporting.
- It does not replace regulation: access, scope and redaction categories are still set by the company.
- It will take time to judge, because the proof is in METR's published reports, not the announcement.
When you read any AI safety claim, ask three questions: who ran the test, who chose what they could see, and who could stop publication. This announcement answers the third differently from most.
01Who is METR?
A nonprofit that evaluates frontier AI models. Per its own post as summarized by FourWeekMBA, it accepts no payment or donations from frontier AI companies.
02How many METR staff get access?
Six, with company hardware, office space and internal tools roughly comparable to Anthropic's own risk-assessment teams (Anthropic-reported).
03Can Anthropic block METR's reports?
Anthropic says it cannot veto publication, but it can require redactions in narrow predefined categories, and METR can say whether it thinks redactions were justified.
04What is off limits?
Customer data, attorney-client privileged information and model architecture details.
05Is this required by law?
No. It is voluntary, and METR itself notes that voluntary regimes limit independence because companies grant the access.
Curious how different assistants behave when you push on safety topics? LumiChats lets you compare 40+ premium AI models side by side in a single subscription, using the same prompt each time.
