Anthropic Embeds METR Team Inside Lab With Veto-Free Publishing Rights

What Happened

Anthropic has granted a six-person team from the Model Evaluation and Threat Research organisation, known as METR, employee-like access to its internal AI systems, with an unconditional right to publish their findings without Anthropic's approval. The arrangement, confirmed in mid-October 2026, makes METR the first outside evaluator embedded inside a major AI laboratory under terms that explicitly prevent the host company from vetoing or delaying the release of results.

Who Is Involved

Anthropic is a San Francisco-based AI safety company and the developer of the Claude family of large language models. METR is an independent organisation focused on evaluating the capabilities and risks of frontier AI systems. The two organisations have had a prior working relationship, with METR conducting external evaluations of Anthropic models, but the new arrangement significantly expands both the depth of access and the independence of the evaluation process.

Terms of the Arrangement

Under the agreement, the six METR staff members receive access comparable to that of Anthropic employees, enabling them to examine model internals, run tests, and review safety-relevant documentation. The key structural feature of the arrangement is the unconditional publication right: METR may release its findings regardless of their content or conclusions, and Anthropic has no contractual mechanism to block, delay, or require modification of those findings before publication. The terms were reported by Tech Times and tech-insider.org, both citing the formal agreement between the two organisations.

Background

The announcement follows a period of intensifying public debate over AI safety practices at frontier laboratories. Anthropic Chief Executive Dario Amodei has previously called publicly for a slowdown in AI development to allow safety research to keep pace, a position that has drawn responses from other AI executives and policymakers. The broader policy environment has also shifted, with safety-focused political action committees gaining ground in recent election cycles, according to reporting by Transformer newsletter, and with growing legislative and legal pressure on AI companies regarding liability for harms caused by their systems.

External or third-party evaluations of AI models are not new. Several laboratories, including Anthropic and OpenAI, have previously commissioned pre-deployment assessments from outside researchers. However, those arrangements have typically involved shorter-term access, narrower scope, and publication terms subject to company review periods or other restrictions.

What It Means in Practice

The structural difference in this arrangement is the absence of a company veto. In prior third-party evaluation frameworks, the laboratory retained the ability to review findings before publication, raise objections, or negotiate the timing and framing of results. Under the Anthropic-METR agreement, METR retains full editorial control over what it publishes and when. This means safety concerns, capability findings, or risk assessments that Anthropic might consider sensitive or reputationally damaging could be published without the company's consent.

The six embedded staff members will operate with access to systems and documentation not available to conventional external researchers, allowing for a more granular evaluation of model behaviour, internal safety mechanisms, and potential failure modes.

Industry Context

The agreement arrives as AI companies face increasing scrutiny from regulators, legislators, and the public over the adequacy of self-reported safety practices. A former SpaceX AI engineer was separately scheduled to testify before the New York City Council in October 2026 on related workplace and safety concerns in the AI sector. Legal pressure on AI companies has also grown, with commentary and editorial voices calling for expanded liability frameworks for AI-related harms.

METR's findings from the embedded evaluation period are expected to be published on a timeline determined by METR, with no announced deadline or scheduled release date from either organisation.