Anthropic brings in METR and SecureBio for outside risk assessments

Anthropic has contracted two outside organizations — the AI evaluation nonprofit METR and the biosafety research group SecureBio — to help assess the risks and safety practices around its models. The company is also in early-stage discussions with other organizations, including Hugging Face, the open-model platform company, according to the announcement.

The move matters because frontier AI labs have largely graded their own homework. External evaluation, where an independent body tests a model before or after release, is one of the few mechanisms regulators and researchers can actually inspect. Which capabilities get tested, and whether results can delay a launch, remains the unresolved question.

Who are METR and SecureBio?

METR (Model Evaluation & Threat Research) is a nonprofit that stress-tests frontier models for dangerous capabilities, such as the ability to autonomously carry out long-horizon tasks — multi-step work a model performs without human intervention. SecureBio focuses on biological risk, an area where policymakers in Washington and Brussels have pressed labs for concrete safeguards.

Anthropic did not disclose the scope, duration, or financial terms of either engagement. The announcement names the partners but not the deliverables. That gap matters: outside evaluation without publication or binding consequences can function as reassurance rather than restraint.

What is Hugging Face's role?

The discussions with Hugging Face are explicitly early-stage, per the announcement. No agreement, timeline, or evaluation framework has been described. Hugging Face hosts open-weight models and has built its own evaluation tooling, so a partnership could plausibly extend Anthropic's external testing infrastructure — but nothing of the sort is confirmed.

Why external evaluation, and why now?

Regulatory pressure is one driver. Both the EU AI Act and voluntary commitments adopted by major US labs call for independent testing of frontier systems. Anthropic, which positions safety as its differentiator in a market dominated by OpenAI and Google, has an interest in demonstrating that its safety claims survive outside scrutiny.

The stakes are also competitive. If METR or SecureBio findings can influence release decisions, that sets a precedent rivals will face pressure to match. If findings stay internal, competitors gain a ready-made excuse to do the same.

What is not said

The announcement leaves several questions open:

• No publication commitment for evaluation results is stated.
• No mechanism is described that would let an outside evaluator block or delay a model launch.
• No date, budget figure, or named model under assessment appears in the announcement.
• The status of the Hugging Face discussions — beyond "early-stage" — is unspecified.

Each of these omissions is standard for a first announcement. Each will also determine whether this becomes a genuine check on the lab or a communications exercise.

The practical test arrives with the first evaluation cycle: either results and constraints become visible to outsiders, or the partnerships remain an assurance without evidence.