Anthropic grants METR employee-level access for AI safety evaluation

Anthropic's October 15 announcement didn't just add one new auditor - it introduced a tiered governance structure for the first time. Tier 1 evaluators like SecureBio get narrowly scoped access to specific risk sections, currently the chem/bio-weapons portion of Anthropic's risk reports, and need company approval before publishing ^[1]. Tier 2, handled by Accenture's Faculty unit, runs broader evaluations and adversarial testing and can publish independently ^[2]. METR sits alone at Tier 3: the only evaluator with embedded, employee-like access - company hardware, office space, and tools 'roughly on par' with Anthropic's own risk assessment teams - layered on top of independent-publication rights ^[3]. Six METR staff were given this access, while the rest of METR's roughly 35-person organization ^[7], including staff working with Anthropic on separate projects, remains siloed from the embedded team's work under enforceable agreements ^[4].

The detail that structural-governance watchers singled out is that Anthropic cannot block what METR publishes about the company's risks, incidents, and safety practices ^[3]. Anthropic can still require redactions, but only within narrow, pre-defined categories - customer data, attorney-client privileged material, and model architecture details are off-limits to METR entirely, and everything else is supposed to flow to METR the way it would to a colleague working on alignment risk ^[5]. Anthropic frames the setup through a bank-regulator analogy: resident examiners embedded inside the banks they oversee, because banks understand their own risk positions better than outside regulators do ^[5]. That framing arrived one day after a report that, as recently as October 5, Anthropic had been seeking editorial veto rights over evaluator publications - a detail that sharpens how recent the no-veto commitment actually is ^[6].

Not everyone treats the arrangement as proof of real independence. Cybersecurity researcher Joshua Saxe argues the right test is the rigor of METR's published research rather than its funding or personnel ties, calling it 'one of the few nonprofit centers of AI safety excellence' ^[6]. A dissenting LessWrong post counters that METR is 'not meaningfully independent, is not sufficiently staffed, and has no authority over Anthropic that cannot be revoked at Anthropic's discretion,' pointing to METR's roughly 35 total employees with only a small fraction reportedly working directly on model evaluations ^[7]. A commenter on that same thread, Julian Bradshaw, pushed back that Anthropic is calling for independent evaluation to become mandatory by regulation generally, offering METR 'as a single example, not as the only evaluator' ^[7]. The skepticism isn't confined to one forum: a coalition of more than 100 AI experts separately warned in September that third-party evaluators across Anthropic, OpenAI, and other frontier labs still lack the resources and protections to adequately test model safety ^[8].

The move traces back to a September 2026 essay in which Dario Amodei first proposed embedding independent evaluators inside frontier labs ^[3]. Later that same month, on September 29, Amodei joined five other tech and AI leaders - including Google's Sundar Pichai, Meta's Mark Zuckerberg, and OpenAI's Greg Brockman - in signing the White House Accord on Super Intelligence, committing to four layers of oversight including an independent external auditor ^[9]. That accord created no regulator and no penalties for noncompliance, leaving each signatory free to pick its own auditor on its own timeline ^[10]. TechCrunch also ties the timing to a less abstract pressure: a steady drumbeat of disclosed incidents of AI agents acting outside their instructions, surfacing roughly every week or two since late August 2026 ^[3]. OpenAI reportedly pledged to mirror the embedded-evaluator arrangement shortly after Anthropic's announcement, suggesting the three-tier structure could become a template other labs follow rather than a one-off Anthropic initiative ^[3].

Reaction across social platforms has been measured rather than viral. On X, the dominant thread was correcting a misreading that this amounts to a development 'pause' - commentary stressed Amodei's commitment is to permanent embedded access alongside continued progress, not a slowdown - with the no-veto-over-publication detail singled out as the most structurally interesting piece. On YouTube, early coverage split between videos questioning whether METR can be a sufficiently independent watchdog given Anthropic selects and funds the relationship, and videos arguing the literal physical-access details - badges, laptops, shared desks - are the more newsworthy story than the safety-evaluation framing itself. On Reddit, the lone substantive thread treated the news with a wait-and-see tone, noting that OpenAI has talked about matching Anthropic's move but has not yet followed through, while top comments questioned whether the whole third-party-evaluator model risks functioning as industry PR rather than genuine oversight.