"This is the best news I’ve seen all week," said John Strand, speaking to a development that Anthropic announced in a company release last Thursday: the company has begun to roll out "employee-like access" to a METR team that will assess its frontier AI models.
Anthropic's "employee-like access" for a six-member METR team
The company announcement says six members of METR, a nonprofit that assesses frontier AI models, will be involved in the assessment. Anthropic described the arrangement as giving the METR researchers "employee-like access." The METR team will conduct its evaluation with that level of access and will publish its risk evaluation.
Publication rights and allowed redactions
Anthropic cannot veto METR's publication of the risk evaluation, the announcement states. However, the company retains the ability to redact information from METR’s public report if that information falls into "predefined categories." The specific categories are not listed in the announcement quoted by the source.
The cyber insurance questionnaire just landed. Now what?
SOC 2, HIPAA, insurance renewals - someone has to own security strategy. Nubivance provides fractional CISO leadership without the full-time salary.
Get a security leadDarin Fredde: access lets researchers "see how you arrived at the answers"
Darin Fredde, Sr. Director of Technical Marketing Engineering at Ridge Security, framed the move as an encouraging step for independent scrutiny. He contrasted surface review with deep access, saying there is "a difference between asking someone to review your homework and letting them see how you arrived at the answers," and argued the latter "offers a much better opportunity to discover what might have been overlooked." From an offensive security perspective, Fredde said independent evaluation brings fresh assumptions and freedom to ask questions internal teams may not have considered. He concluded that independent testing "isn't about proving someone wrong. It's about building confidence through evidence rather than assumptions."
Jacob Krell: "Six people cannot audit a frontier AI lab"
Jacob Krell, Senior Director: Secure AI Solutions & Cybersecurity at Suzu Labs, was more skeptical about whether the access as described amounts to meaningful audit capacity. "Six people cannot audit a frontier AI lab," he said, calling the employee-level access "a real step" but cautioning that the bank examiner analogy used by Anthropic cuts the other way. Krell noted that large banks receive teams of resident examiners, rotating staff, and decades of examined process, while six embedded researchers will only "get one slice of one company's output."
Krell warned that "employee-level access means employee-level volume" — internal risk reviews, training run documentation, incident reports, and eval results produce a large "firehose" of material. He said a six-person team "reading that firehose while also designing its own evaluations will spend most of its capacity just keeping up," and that "something goes unreviewed, and the things that go unreviewed are never the things a press release mentions." Drawing on past audit patterns, Krell said two auditors can find genuine issues in the areas they reach while the remainder becomes "sampling error dressed up as assurance." He argued the real test will be "whether Anthropic scales embedded headcount with model capability," calling the current arrangement "a pilot" and urging readers to "treat the findings accordingly."
John Strand: voluntary employee-level access as governance and business risk mitigation
John Strand, Owner of Black Hills Information Security, welcomed the move as an example of voluntary oversight rarely seen in the field. He said it is "exactly the type of oversight and transparency people have been asking for" and expressed hope that it "becomes standard across the companies building frontier models." Strand argued some companies fear oversight will slow innovation, but cautioned they are "forgetting something pretty basic: their AI models going rogue is an actual business risk." In his view, "good governance means doing more than moving as fast as possible and maximizing profits at every turn," and responsible oversight is "part of keeping a company around for the long term."
Anthropic's announcement establishes a clear framework: METR receives employee-like access, six researchers will conduct the assessment, METR may publish its findings, and Anthropic may redact items in specified categories. Security leaders quoted in the source split between endorsement of the transparency and concern that six embedded researchers may be insufficient to exhaustively audit a frontier lab. Whether that balance—access versus capacity—yields a substantive, broadly convincing risk evaluation will be the central question as METR completes and publishes its work.
https://www.securitymagazine.com/articles/102665-anthropic-gives-metr-team-employee-like-access
