Summary: On October 15, 2026, Anthropic announced that six METR employees now have employee-like access — company hardware, office space, and internal tools roughly on par with its own risk assessment teams — and METR will publish its risk assessments without Anthropic veto. For builders, this is a concrete answer to a hard question: what does meaningful third-party oversight of frontier models actually look like when most safety information can’t be shared publicly?
The problem independent evaluators solve
Anthropic’s argument is straightforward: as models get more capable, safety decisions matter more, but the information needed to judge those decisions — training details, compute, how safeguards function — is either commercially sensitive or dangerous to publish. So outsiders can’t verify a company’s risk claims on their own.
The fix Anthropic describes is borrowed, loosely, from bank supervision. Regulators place resident examiners inside the largest banks because banks know their own risk positions better than anyone outside. The company names three things an evaluator needs to make this work: frontier-model expertise, independence (no equity or financial dependence), and enough access plus freedom to publish. Anthropic adds a fourth practical note — evaluators must understand how evaluations themselves go wrong, citing alignment faking as an example.
What “employee-like access” actually means
The specifics matter more than the label. Six METR staff are onboarded with Anthropic hardware and internal systems. Other employees are encouraged to share information with them as they would with a colleague on alignment risk. A short list of carve-outs applies: customer data, attorney-client material, and model architecture details.
Publication rights are the interesting part. METR can publish its analysis of Anthropic’s risk levels, incidents, and practices, and Anthropic cannot veto publication — it can only require redaction within the narrow pre-defined categories. METR is then free to publicly argue whether those redactions were justified, as long as it doesn’t reveal what was redacted.
There are structural guardrails too. The engagement is limited to a small team under enforceable agreements, and other METR researchers — including ones doing other work with Anthropic — are siloed from it. Anthropic frames this as embedding a few qualified people under its security standards, not an information channel for other organizations.
A three-tier ladder for everyone else
METR sits at tier three. SecureBio, currently reviewing the chemical and biological weapons chapter of Anthropic’s August Risk Report, is at tier one: narrowly scoped investigations with publication requiring approval. Tier two allows publication without approval plus broader access, still subject to category-based redactions. Accenture and Faculty are part of a broader partnership covering red-teaming, alignment assessments, and safeguard testing. Anthropic says early-stage partners get less access and fewer publication rights until trust is built, and that it’s in early discussions with several others, including Hugging Face.
Why this matters if you build on these models
Two takeaways for practitioners. First, evaluation access is becoming a differentiator you can watch. If independent teams start publishing evidence-based assessments of frontier labs’ risk posture, you get a signal about model safety that doesn’t depend on the lab’s own Risk Report. This continues a thread this blog has followed: Anthropic’s earlier Accenture deal introduced embedded evaluation as a concept, and this update shows what the deepest tier of that looks like in practice (Embedded Evaluators: What Anthropic’s Accenture Deal Changes for Builders).
Second, the tiered structure is a design worth studying if your own product needs external audits — security, compliance, or model behavior. Anthropic’s progression from scoped investigations to full embedding, with publication rights expanding as trust accrues, is a reusable template for any vendor relationship where confidential data and public accountability pull in opposite directions.
The honest caveat: this is voluntary, single-company, and early. METR’s first published assessment under these terms will tell us more than the framework description does. Until then, treat this as promising infrastructure, not proven oversight — and watch for whether other frontier labs follow Anthropic’s call to embed evaluators too.