Anthropic said on October 15, 2026 that six researchers from the nonprofit Model Evaluation and Threat Research, or METR, now have employee-like access inside the company to evaluate advanced AI models, safety practices and risk management. The arrangement gives outside evaluators company hardware, office space and access to internal tools roughly comparable to what Anthropic’s own risk-assessment teams use.

In Anthropic’s announcement, the company said METR will be allowed to publish its own assessments without Anthropic having a general veto over publication. Anthropic can still require narrow redactions for predefined categories such as customer information, attorney-client privileged material and sensitive model-architecture details. METR may publicly say whether it believes a requested redaction was justified, provided it does not reveal the protected information itself.

We’ve committed to having evaluators assess our safety practices. We're in discussions with multiple organizations, and have now given employee-like access to a team from @METR_Evals.

In this post, we describe how we work with third-party evaluators: https://t.co/8CLQnMVg4X

— Anthropic (@AnthropicAI) October 15, 2026

The access is deep, but not unlimited

Anthropic describes three levels of evaluator access. METR is currently in the highest tier, where evaluators receive the deepest internal access and retain independent publication rights. SecureBio is in the first tier, conducting more narrowly scoped reviews, while Anthropic says it is also working with Accenture and Faculty and has held early discussions with additional groups.

Reuters reported that the program arrives amid wider concern over how much control AI companies can exert over outside reviewers through access restrictions and publication rules. The new structure does not make METR a government regulator, and the arrangement remains voluntary. Its value will therefore depend heavily on what METR is able to inspect, what it ultimately publishes and how often Anthropic requests redactions.

Why the timing matters

The move follows a string of concerns about increasingly capable AI agents operating outside expected boundaries. BitcoinVersus.Tech recently reported that Anthropic cut internet access during internal evaluations after unintended model actions. We also covered DeepMind’s warning that recent AI-agent security incidents should be treated as early warning shots.

Anthropic compares the METR setup to resident examiners inside large banks: outsiders need enough access to understand risks that are difficult to see from public information alone. That comparison is useful, but there is an important difference. Bank examiners operate under formal regulatory authority; METR’s access exists because Anthropic voluntarily granted it.

The broader industry question is whether this becomes a repeatable model. The White House AI accord already pushed major developers toward independent review. Anthropic’s latest step puts more detail behind that commitment by defining who gets access, what can be published and what remains protected. The next test will be METR’s first report produced under this arrangement.

---

BitcoinVersus.Tech Editor’s Note:
We volunteer daily to ensure the credibility of the information on this platform is Verifiably True. If you would like to support to help further secure the integrity of our research initiatives, please donate using the QR code provided

For best QR-code scanning results, please scan this donation code from a desktop or larger screen. Mobile screens may prevent camera recognition.