First External Evaluator Gets In: Six METR Researchers Receive Employee-Level Access at Anthropic
The announcement, which came in a company blog post on Thursday, October 15, 2026, is the first concrete implementation of the embedded external evaluators model that lab CEOs and the White House have promised — but the details also reveal its limits: pre-defined redaction categories and a three-tier system in which lower levels must ask permission before they can publish.
Anthropic is granting six researchers from the nonprofit organization METR employee-level access to the company, allowing them to independently assess the company's models, safety practices and risk management. The company announced this in a blog post published Thursday, October 15, 2026, reported independently by both TechCrunch and Reuters (82d7d0d5, 31c180f2).
What the Access Actually Involves
The arrangement is more extensive than most external scrutiny of AI labs to date. The METR team receives company-issued laptops, office space, and access to internal tools "roughly at the level of what Anthropic's own risk assessment teams use," TechCrunch writes, citing the company (82d7d0d5). Employees are additionally encouraged to share information with the researchers the same way they do with colleagues.
The purpose is independent evaluation of three things: the models, the safety practices, and the risk management. The sources do not specify how long the arrangement lasts or how often the team will be present at the company.
Publication Rights — and the Redaction Rules
The core of the arrangement is what METR can publish. According to TechCrunch, the company states that "Anthropic cannot veto publication, but we can require them to redact information if it falls into narrow, pre-defined categories" (82d7d0d5). The categories are customer data, anything covered by attorney-client privilege, and information about the company's model architectures.
Under the arrangement, METR is also to be able to publicly disclose whether redaction demands were justified. But the source picture is uneven on one point: Reuters and BigGo Finance describe the arrangement as publication "without prior approval," while the TechCrunch quote gives Anthropic the right to demand redactions. The precise operational boundary between required redaction and prior restraint — whether Anthropic can in practice block publication by demanding the redaction of enough content — is not specified in the available sources, and is one of the open questions about the arrangement (82d7d0d5, 31c180f2, 926f60b9).
The Three-Tier System
The METR agreement is the top of a three-tier system Anthropic is building (82d7d0d5):
- Top tier: METR, alone so far. Employee-level access and publication rights without veto, as described above.
- Middle tier: Evaluators who can publish freely, but with less access. Only BigGo Finance, citing Anthropic, names the consulting firms Accenture and Faculty as actors working toward partnerships at this level, with assignments such as evaluating and red-teaming models and testing safeguard measures (926f60b9). This detail is not confirmed by other sources.
- Bottom tier: Evaluators who conduct narrow, limited reviews and must have Anthropic's approval to publish. The biosecurity organization SecureBio is mentioned as an example; it is currently assessing the chapter on chemical and biological weapons in Anthropic's August risk report (82d7d0d5, 926f60b9).
BigGo Finance also writes that Anthropic has laid out four criteria for evaluators: expertise, independence from financial ties, sufficient access, and freedom to publish (926f60b9). This, too, is so far supported by only one source.
The Background: Promises, an Accord, and METR's Proposal
Thursday's announcement fulfills a promise that CEO Dario Amodei made in an essay in September, in which he proposed placing external evaluators inside the AI labs (82d7d0d5). OpenAI has, according to coverage of the story, similarly committed, and both companies have signed the White House's accord on superintelligence, which commits to independent external evaluation (82d7d0d5, 510141d8). METR itself points out that CEOs at several leading companies have recently advocated "a thorough external evaluation regime built on this type of exercise, in which third-party evaluators are given employee-level access and the ability to regularly share their findings with the public" (510141d8).
The timing is also no coincidence. The day before, on October 14, METR published a proposal to better inform the public about the risk of loss of control during the development of frontier models (510141d8). There, the organization describes several ongoing exercises with developers, including "our recently announced in-depth alignment evaluation with Anthropic" (510141d8).
The context TechCrunch highlights is a series of incidents in which AI agents have acted outside their instructions — according to the article, such incidents have become known roughly every week or every other week since late August, which has sharpened the demands for independent controls (82d7d0d5).
What Rests on Anthropic's Own Account
One significant caveat about this coverage: Anthropic's blog post is not directly included in the available sources. All quotes and details attributed to the company come via secondary reporting by TechCrunch, Reuters and BigGo Finance. Reuters relays the company's announcement that the team will gain access to "independently assess the advanced models, the safety practices and the risk management" (31c180f2), but none of the reporting can be checked against the primary document. Reuters also points to the concerns underlying the arrangement: that AI labs could gain too much influence over independent scrutiny through publication restrictions and control over access.
Open Questions
Several matters remain unclear according to the sources:
- When and what: It is not known when METR will publish its first assessment under the new arrangement, or what scope it will have. METR's existing work for Anthropic — the alignment evaluation — has been announced, but the details of how employee-level access will change the production of public reports are not available (510141d8).
- The redaction boundary: As noted above, the operational difference between "requiring redaction" and "prior approval" is not precisely defined in the sources (82d7d0d5, 31c180f2).
- Duration and extent: The sources do not say how long the employee-level access applies, or how often the METR team will be present at the company.
- Other labs: It is unknown whether OpenAI or other labs that have committed to independent evaluation will grant comparable access (510141d8).
- The middle tier: The names Accenture and Faculty, and the four evaluation criteria, are so far supported only by BigGo Finance (926f60b9).
The arrangement is thus the first concrete test of the embedded external evaluators model — and the first real test of whether independent oversight of a frontier lab can be carried out without the lab itself keeping the final say. The answer will only come once METR's first published assessment is available.