Anthropic lets outside auditors publish. Almost.

On October 15, Anthropic handed a nonprofit safety group the one thing evaluation talks usually fight over: the right to publish without asking. A six-person team from METR (Model Evaluation and Threat Research) gets employee-like access to Anthropic's systems, company hardware, office space, and internal tools roughly equivalent to what the company's own risk-assessment teams use. METR will assess Anthropic's risk levels, incidents, and practices and release its findings without Anthropic having a general veto.

The access is the headline. The publishing terms are the story. Anthropic can require narrow redactions, but only for predefined categories: customer data, attorney-client privileged material, and model-architecture details. And METR gets to say publicly whether it believes a requested redaction was justified, without revealing the protected information itself. The default is disclosure, not approval. That is a genuine concession, and it matters that it follows pressure: Reuters reported exclusively earlier in the week that Anthropic and OpenAI had been seeking significant control over what outside reviewers could publish. A person familiar with the talks told Reuters that Anthropic had grown more open in recent negotiations.

The arrangement sits inside a new three-tier evaluator framework. SecureBio, the biosecurity nonprofit, operates at tier one and is currently reviewing Anthropic's August 2026 risk report, focused on the chemical and biological weapons section. It previously assessed the February 2026 report, rating the Claude Opus 4.6 model "very low but not negligible" for existing threats and "low but uncertain" for novel risks. Accenture's Faculty unit handles tier two: broader evaluations and adversarial testing. METR gets the most hands-on access. Anthropic is reportedly also in talks with Hugging Face, which is unconfirmed.

Two caveats keep this from being the whole answer. First, no METR evaluation has been published from this access yet, so the framework's value is prospective. A redaction regime this narrow is only as strong as the first time someone tries to stretch it. The line between protecting dangerous information and protecting a company's reputation can blur, and the history of "independent" evaluation funded and arranged by the lab itself has reasons for skepticism. Second, Anthropic chooses its evaluators. Independence arranged by the evaluated party is still partial, which Reuters' sources also flagged.

For operators choosing vendors, the signal is narrower than the announcement. The meaningful commitment is publication without veto, with a redaction-dispute mechanism METR can invoke publicly. That is a better disclosure regime than the industry's current default, and it raises the bar for what competitors offer their own reviewers. But nothing in the announcement changes what you can observe about a deployed model from the outside. Treat it as governance upside on Anthropic's ledger, not as an independent safety verdict on Claude.

What remains unknown: what METR actually publishes, how the redaction clause behaves under dispute, and whether OpenAI matches the arrangement. Anthropic's own blog post on independent evaluators was not opened directly for this briefing; the quotes above come via its October 15 announcement as reported.

Related: the FTC agent-safety probe is the regulatory side of this same scrutiny, and the eight-state AG suits against OpenAI cover the enforcement angle from yesterday's edition.