Anthropic says six METR researchers can work with access roughly comparable to its own risk-assessment teams and publish their findings without a company veto. The company also retains the final say over whether information falls into categories that must be redacted.

Those two provisions shape what the public can learn from the arrangement. Broader access could help an outside evaluator investigate risks that ordinary model testing misses. Publication rights let it report conclusions without seeking approval. Redaction authority still gives the company influence over the evidence readers can inspect.

Anthropic’s October 15 announcement, clarified on October 16, describes the arrangement as voluntary. It gives an external evaluator a closer view of the company’s work; it does not create statutory supervision or certify that a model is safe.

Anthropic’s three tiers separate access from permission to publish

METR is an external AI evaluation organization conducting an in-depth alignment assessment with Anthropic, an engagement METR itself confirms. Alignment assessments examine whether AI systems might act against human intentions, including risks of losing control over advanced systems.

Anthropic places the embedded METR arrangement in the third tier of its developing evaluator framework. Its announcement describes materially different rights across the three tiers:

Tier Access described by Anthropic Publication and redaction terms described
1 Narrow investigations, with Anthropic deciding access Publication requires Anthropic’s approval. The tier description does not specify an additional general disclosure right.
2 More information Evaluators can publish without approval. Anthropic can require removal of specific categories of information; evaluators can publish their views on whether redactions were justified.
3 METR’s employee-like access Publication rights are similar to tier 2. The METR-specific description adds that Anthropic ultimately decides category membership, while METR reserves certain statements about the access and information it received.

These are Anthropic’s public descriptions, rather than a review of the underlying agreements. The announcement does not assign numbered tiers to SecureBio, Accenture or Faculty, so METR’s terms should not be assumed to apply to those partners.

For METR, “employee-like” means company hardware, office space and access to internal tools and systems “roughly on par with what our own risk assessment teams have access to,” according to Anthropic. Staff are encouraged to share information as they would with an alignment-risk colleague, outside defined exclusions.

Customer data, attorney-client privileged information and model architectures are excluded from that access. Anthropic qualifies the architecture exclusion: it says it will generally seek to share highly risk-relevant architecture information with explicit internal approvals. These access limits concern what researchers can see during their work.

Anthropic retains the final say over redaction categories

Publication redactions concern what researchers can subsequently tell readers. Anthropic says it can require METR to remove information in predefined categories, giving reproduction of confidential intellectual property as an example.

The consequential wording is explicit: “Anthropic ultimately decides what information does and does not fall into these categories.” That leaves the company with substantive control over parts of a report even though it cannot veto publication as a whole.

The same sentence says METR reserves the right to make “certain statements about the level of access and information it received.” This gives readers a potential way to understand constraints on the assessment. The announcement does not enumerate every reserved statement or every redaction category, and it does not promise unrestricted disclosure of withheld material.

The distinction matters when comparing engagements. Our coverage of METR and Redwood’s OpenAI monitoring assessment describes an earlier pilot with company redaction rights and no unilateral right to describe redactions. Publication and disclosure rights need to be read for each engagement; they do not follow automatically from an evaluator’s name.

METR still depends on companies granting access

The Financial Times, in reporting published by Ars Technica on October 18, reported that Anthropic is not paying METR for this arrangement.

That is consistent with METR’s October 14 account of its funding approach. METR says it has so far chosen not to accept payment or donations from frontier AI companies. It also accepts substantial free model access, while saying it could afford those tokens without materially affecting its operations. The general policy provides context; it is not an independent financial audit of this engagement.

METR identifies a separate dependence in the same policy: its work relies on companies voluntarily granting access and permission to publish. Information supplied by a company can also be difficult to verify and may be selectively filtered. An evaluator can avoid being paid by the subject of its work while still needing that subject’s cooperation to investigate it.

METR’s transparency policy seeks disclosure of consequential access limits, redactions, publication restrictions and informal pressures. It expressly cautions that not all previous or in-flight work meets that standard. Those commitments therefore cannot stand in for the specific Anthropic agreement.

Anthropic says the embedded team works under enforceable agreements and is siloed from other METR staff, including researchers on other Anthropic projects. METR’s February confidentiality policy describes project-specific separation of sensitive information. It helps explain how an external organization can handle confidential access, but does not establish this engagement’s publication rights.

METR’s October 20 criticism came from an earlier arrangement

There is already a public example of METR disagreeing with Anthropic over redactions. Its October 20 review of Anthropic’s August risk report says some information withheld from that underlying report was material to METR’s conclusions about particular risk pathways.

The review’s date can be misleading. METR explicitly says the work began before the evaluator announcement, used a relatively lightweight review of Anthropic’s internal report, and did not have the newly described access. Anthropic could veto publication or redact any part of that review under the earlier terms.

METR says Anthropic made no redactions to METR’s review and its conclusions did not change during publication review. Its criticism concerned redactions in Anthropic’s underlying report. METR judged some of those redactions unjustified and said withheld information reduced its confidence in monitoring and raised its assessment of some risks, without changing its overall conclusion.

That is evidence of published disagreement under the earlier terms. It cannot demonstrate how the newly announced third-tier arrangement handles a dispute.

The agreement’s operation remains an open question

The public sources reviewed as of November 4 do not establish how the newly announced terms have worked in practice. They also do not supply the complete agreement, specific redaction deadlines, engagement duration, termination protections, leadership or legal-access clauses, or a dispute-resolution mechanism. Those terms remain unknown from the inspected material.

For future assessments, useful evidence would include what access METR actually obtained, which consequential limitations it could disclose, and how a contested redaction affected the published account.