Anthropic is letting outside safety researchers into the building. Literally.
On October 15, 2026, the AI developer expanded its work with independent evaluators. Under the expansion, a six-member team from the nonprofit METR (Model Evaluation and Threat Research) receives what amounts to employee-like access to the company’s hardware, office space and internal systems. The biosecurity nonprofit SecureBio and Accenture’s Faculty unit also hold defined roles in a new three-tier evaluator framework.
The interesting part is not the badge access. It is the publishing rights. METR can disclose what it finds without asking Anthropic for permission first. For a company building frontier AI models, that is a notable amount of rope to hand an outside group.
How the three tiers work
SecureBio operates at tier one. The nonprofit is currently reviewing Anthropic’s August 2026 risk report. It is focusing specifically on the section covering chemical and biological weapons.
This is not SecureBio’s first pass at Anthropic’s homework. It previously evaluated the company’s February 2026 risk report. In that earlier assessment, SecureBio judged the risks tied to the Claude Opus 4.6 model as “very low but not negligible” for existing threats. For novel risks, it landed on “low but uncertain.”
Accenture’s Faculty unit handles tier two. Its remit covers broader evaluations and adversarial testing.
METR rounds out the arrangement with the most hands-on access of the three groups described. Its six researchers can work directly with company hardware and internal systems. They are positioned to evaluate and publish findings on risks associated with Anthropic’s Claude models.
There is a caveat on the publishing side. Anthropic may make limited redactions for sensitive information. So METR’s reports won’t be entirely unfiltered. Still, the default is disclosure, not approval.
Why independent evaluation matters here
The arrangement ties back to commitments in Anthropic’s Advanced AI Framework. The company has long positioned itself as the safety-focused player among frontier AI developers.
Anthropic is also reportedly exploring potential collaborations with Hugging Face. That would further widen its network of safety reviewers. No such partnership has been confirmed.
What this means
The framework is only as meaningful as what comes out of it. As of the announcement date, METR has published no evaluation results from its new access.
The redaction clause deserves close attention. “Limited redactions for sensitive information” is a reasonable carve-out. But the line between protecting dangerous information and protecting a company’s reputation can blur.
Biosecurity remains the area to watch most closely. SecureBio’s review of the August 2026 report will show whether its earlier “very low but not negligible” and “low but uncertain” assessments hold for newer work.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.