Anthropic has granted a team of six researchers from METR, a nonprofit organization focused on evaluating frontier artificial intelligence systems, an unprecedented level of access to its internal operations. The researchers will independently assess Anthropic's AI models, safety practices, and risk management processes, the company announced in a blog post published Thursday.

The move comes during a period of growing scrutiny over AI agent behavior. Since late August, incidents involving AI agents operating outside their intended instructions have been surfacing roughly every one to two weeks. Those disclosures have intensified pressure on AI developers to submit to independent oversight rather than relying solely on internal checks.

"We want to be publicly accountable for the decisions we're making. We aim to close the gap between what evaluators understand about a frontier AI company's systems and what the company itself understands."

Employee-Level Access for Outside Researchers

Under the terms of the arrangement, the METR team has been provided with Anthropic-issued laptops, dedicated office space, and access to internal tools that are broadly comparable to those used by Anthropic's own risk assessment personnel. According to the company, its employees are actively encouraged to share information with the METR researchers in the same way they would with a colleague working on alignment risk.

Anthropic drew a parallel between this arrangement and the practice in financial regulation, where regulators embed resident examiners inside major banks to maintain ongoing visibility into their operations. The analogy reflects the company's framing of this as a structural accountability measure rather than a one-off review.

Publication Rights and Limits on Veto Power

One of the more significant aspects of the arrangement is that Anthropic has relinquished the right to block METR from publishing its findings. The nonprofit will independently produce and release its own assessments of Anthropic's risks, incidents, and operational practices without requiring prior approval from the company.

Anthropic does retain a narrowly defined ability to request redactions. Specifically, the company can require that certain categories of sensitive information be removed before publication, including customer data, attorney-client privileged communications, and details relating to model architectures. However, the scope of that authority is explicitly constrained.

"METR is free to characterize our redactions publicly, and say whether it believes the redactions were justified, as long as it does so without revealing the redacted information," Anthropic stated. This provision ensures that even the act of redacting information remains subject to a form of public accountability.

A Three-Tier Oversight Structure

Anthropic is building a tiered framework for external evaluators, and METR currently occupies the top level of that structure. The system is organized into three distinct tiers, each offering a different degree of access and publication autonomy.

• The bottom tier includes evaluators such as biosecurity nonprofit SecureBio, which conduct narrow, focused reviews and must obtain Anthropic's approval before publishing their findings.
• The middle tier provides evaluators with a broader set of information and allows them to publish without seeking prior approval from Anthropic.
• The top tier, currently occupied exclusively by METR, adds the full employee-level access described in Thursday's announcement.

The structure is designed to allow different types of evaluators to engage with Anthropic's systems at varying depths, depending on their scope and expertise.

Fulfilling a Public Commitment

Thursday's announcement follows through on a pledge made by Anthropic CEO Dario Amodei in a widely circulated essay published in September. In that piece, Amodei proposed embedding independent evaluators inside AI laboratories as an early and concrete step toward introducing meaningful checks on the pace of AI development.

Shortly after that essay appeared, OpenAI publicly committed to pursuing a similar approach. Since then, both Anthropic and OpenAI, along with Google, Meta, Nvidia, and xAI, have signed a White House accord that formally commits each company to independent review of their AI systems.

The agreement with METR represents one of the more concrete implementations of that commitment to date, establishing a working model for what ongoing, embedded external oversight of a frontier AI company might look like in practice.