Announced 15 Oct 2026 · Sources checked
What Anthropic announced on 15 October
Anthropic published An update on independent evaluators on 15 October 2026. The headline commitment is concrete: the company has begun giving "employee-like access" to a team from METR, and METR will publish the results of its risk assessments.
The post links the move to two earlier promises. Last month Anthropic said it would strengthen independent third-party assessment of its models and safety practices. It also signed the White House Accord on Super Intelligence, which asked companies to partner with independent external auditors or evaluators. Today's post is the first named embedding under that frame.
TechCrunch's same-day report places the announcement against a run of disclosed agent incidents since late August and notes that calls for independent checks have grown louder. The primary facts below come from Anthropic's own post; secondary coverage is used only for context.
What "employee-like access" actually includes
Employee-like access, as Anthropic defines it, is not a remote API key. The company says it has onboarded six METR employees with company hardware, office space and access to internal tools and systems that are roughly on par with what its own risk-assessment teams have. Outside a short list of carve-outs, Anthropic staff are encouraged to share information with them as they would with a colleague working on alignment risk.
The carve-outs named in the post are data about customers, attorney-client privileged information, and information about model architectures. A footnote explains the architecture caution: because evaluators may also work with other companies, Anthropic worries about inadvertent leaks, and it says architecture details are usually low relevance for risk assessment. When they are highly relevant, Anthropic says it will generally seek to share them after explicit internal approvals.
The arrangement is intentionally small. Anthropic says the engagement is limited to a security-vetted team under enforceable agreements, and that other METR employees are siloed from that team's work, including researchers who work with Anthropic on other projects. Embedded evaluation, in this telling, is a close view for a few people under Anthropic's security standards — not a pipe that lets other organisations obtain Anthropic or customer information.
Publication rights and what Anthropic can still redact
Publication is the other half of the deal. Anthropic says it cannot veto METR's analysis of risk levels, incidents or practices. It can require redactions when information falls into the narrow predefined categories above. METR is free to say in public whether it believes those redactions were justified, as long as it does not reveal the redacted material itself.
That is a meaningful shift from arrangements where a lab can block a report entirely. It is also not full sunshine. Redaction power still sits with the lab for categories it defined, and outsiders will have to judge METR's characterisations of those redactions without seeing the underlying text. The post's bank-examiner analogy is useful here: resident supervisors see more than the public, and the public sees a mediated version of what they found.
Independence criteria in the post are explicit. Evaluators should not hold equity in the company they evaluate or be financially or operationally dependent on it. They need expertise in frontier evaluations, including failure modes such as alignment faking, enough access to see training and safeguards, and freedom to publish. Anthropic presents the METR embedding as an attempt to meet those four conditions at once.
Three tiers of evaluator access
Anthropic says it is developing three tiers of evaluator access and hopes, over time, to have several evaluators at the third tier, each publishing their own view of its risk posture and key decisions.
Tier one is narrowly scoped investigations. Anthropic decides what access to provide, and evaluators publish with Anthropic's approval. Tier two lets evaluators publish without approval and grants more information, while Anthropic can still require removal of specific categories; evaluators may publish whether those redactions seemed justified. Tier three is the METR arrangement: employee-like access plus publication rights similar to tier two.
The tier ladder is as important as the METR headline. It admits that most partners are not yet at the deepest access, and it gives readers a way to score future announcements. A press release that only says "we work with evaluators" is not the same as a named tier-three embedding with public analysis rights.
SecureBio, Accenture and other partners
SecureBio is currently in the first tier. Anthropic says it is reviewing the chemical and biological weapons chapter of the August Risk Report after completing and publishing a pilot review of a previous risk report. After the current review, Anthropic says it will consider projects with more access and stronger publication rights.
With Accenture and Faculty, Anthropic describes an ambitious partnership still being built: evaluating and red-teaming models, alignment assessments, testing safeguards and capacity-building across the organisations. Early discussions are also under way with several other organisations, including Hugging Face. Those partners currently have more limited information and more limited publication rights while trust and process are established.
Anthropic closes with a normative claim: it is embedding evaluators voluntarily, and it believes every frontier AI company should do the same. That is advocacy, not a regulatory requirement, and other labs have not matched this specific METR embedding in a public post we reviewed for this article.
How this relates to earlier fights over evaluator independence
The announcement lands two days after reporting that Anthropic and OpenAI had sought control over when outside safety evaluators could publish. That earlier story, based on Reuters reporting, described contractual fights over publication. Today's Anthropic post does not mention those contracts by name, but it does stake out a different public position: no veto, only narrow redactions, with the evaluator free to criticise redactions.
It also arrives in the same news cycle as FTC civil investigative demands to OpenAI and Anthropic over safety incidents and safeguarding practices. An embedded nonprofit evaluator is not a substitute for a government investigation, and Anthropic does not claim it is. Readers should keep the channels separate: METR's future public notes will be one evidence stream; FTC process, if it becomes public, will be another.
UK policy work is moving in parallel. The UK AI Security Institute's technical best-practice priorities emphasise external evaluation and incident handling for frontier developers. Anthropic's tier language gives a commercial lab's version of what deep access can look like when it is voluntary rather than imposed.
What METR has done with Anthropic before
METR is not new to Anthropic's paperwork. The post says that over the past year the two organisations have worked on several reviews of Anthropic's Risk Reports. The novelty is the depth of access and the publication rules attached to a resident team, not the existence of any METR relationship.
That history cuts two ways. Continuity means METR already knows Anthropic's report formats and past findings, which should speed meaningful work. Continuity also means critics will ask whether familiarity softens independence. Anthropic's answer in the post is structural: no equity, siloed teams, enforceable agreements and a public commitment that it cannot veto METR's analyses.
Until METR publishes under the new rules, those safeguards are promises. The first public assessment that cites employee-like access — and any note about redactions — will be the real test of whether the arrangement changes what outsiders can know.
Limits, caveats and what to watch next
Several limits are built into the announcement. Only six people are embedded. Architecture information is restricted by default. Other METR staff are siloed. Other evaluator partners remain on shallower tiers. Anthropic wrote the redaction categories. None of that invalidates the step; all of it bounds what the step can deliver.
There is also no timeline in the post for METR's first public write-up under the new access, no list of which internal systems are in scope, and no third-party confirmation from METR beyond Anthropic's description in the materials we read for this article. TechCrunch summarises the same Anthropic claims; it does not add a separate METR statement with new operational detail.
What to watch next is therefore specific: METR publications that mention the embedded team; whether SecureBio moves up a tier after the chem-bio chapter review; whether Accenture or Faculty reach tier-two publication rights; and whether other frontier labs announce comparable resident-evaluator deals rather than one-off paper reviews.
Common questions
Can Anthropic block a METR report?
Anthropic says it cannot veto METR's analyses. It can require redactions only for narrow predefined categories such as customer data, privileged legal information and certain architecture details. METR may publicly say whether it thinks a redaction was justified.
Does this give all of METR access to Anthropic systems?
No. Anthropic says only a small, security-vetted team is embedded, under enforceable agreements, and that other METR employees are siloed from that team's work.
Who else is evaluating Anthropic under this programme?
SecureBio is in the first tier, reviewing part of the August Risk Report. Accenture and Faculty are in earlier-stage partnerships with more limited access and publication rights. Anthropic also says it is in early discussions with others, including Hugging Face.
Sources & further reading
How this story was made
Written by Kristian Kostov with AI assistance and checked against the linked sources. Company performance claims are attributed to the company. Analysis reflects AiLookout’s interpretation; we have not independently tested the products discussed. The cover places the companies’ own logos, taken from public-domain or openly licensed files credited in the caption, on a plain background. It is not a product photograph, a screenshot, or an endorsement.
Our editorial standards
