Anthropic gives third-party auditors ’employee-like access’ in AI industry first

Anthropic announced on Thursday that it will give "employee-like access" to six evaluators from METR, a third-party nonprofit that tests frontier AI models. The evaluators have been issued Anthropic hardware and office space, and can reach internal tools and systems that the company describes as roughly on par with what its own risk assessment teams use.
The arrangement is unusual in how far the access extends. Anthropic said that outside a small number of clearly defined categories —customer data, attorney-client privileged information, and information about model architectures are the examples it gave — its employees are encouraged to share information with the METR evaluators much as they would with a colleague working on alignment risk. In other words, the outside testers are being treated less like visiting auditors and more like staff embedded in the safety organization.
METR will publish its findings, and Anthropic is not permitted to withhold any of them, though the company will ask METR to redact sensitive information where necessary. Anthropic framed the arrangement as a way to reconcile two competing needs: being accountable while not publicly sharing all information relevant to its safety work, some of which it says could be useful to malicious actors or is commercially sensitive, such as compute or training information. The company said it aims to close the gap between what evaluators understand about a frontier AI company's systems and what the company itself understands.
Anthropic also gave a rationale for choosing an outside organization at all. It noted that researchers like METR are valuable to the review process because they do not hold equity in the company they are evaluating and are not financially or operationally dependent on it. That independence argument is the crux of the case for third-party evaluation: an auditor whose funding, employment or returns are tied to the lab has an incentive problem that no amount of internal goodwill fully removes.
The idea traces back to Anthropic CEO Dario Amodei, who in a September blog post proposed embedding evaluators from nonprofits like METR as one of three suggestions for "pacing the frontier." The Thursday announcement effectively puts the first of those suggestions into practice.
Context for developers and product teams: METR, formerly the Model Evaluation and Threat Research project, is a nonprofit known for running structured evaluations of frontier models, including agentic and autonomy-related capabilities. "Employee-like access" matters because black-box testing —asking a model questions through an API and scoring the answers —only exercises the surface of a system. It cannot easily inspect training data provenance, internal monitoring, deployment safeguards or the compute budget behind a run. Giving evaluators internal tools and systems is intended to let them see the machinery, not just the outputs, which is roughly the difference between auditing a company's financial statements and being allowed into its accounting system.
The limits Anthropic carved out are worth noting. Customer data, privileged legal information and model architecture details stay off-limits, which means the deepest technical secrets about how the models are built remain inside the company. METR gets broad visibility into risk and safety operations and access to the models themselves, but not a full view of architecture. This suggests the disclosure is calibrated: enough to make safety claims checkable, not enough to hand competitors a blueprint. A likely trade-off is that outside reviewers may still be unable to independently verify claims that depend on architecture details they cannot see.
The announcement lands amid heavier scrutiny of AI labs. In September, OpenAI expressed at least partial support for the FRONTIER Act, a bipartisan bill introduced in July that would require certain AI companies to allow "independent verification organizations," or IVOs, to review the safety of their development process. That was the first time OpenAI had backed proposed federal regulation of AI, though it has at times favored unified federal rules over a patchwork of state AI laws.
Federal regulation under the Trump administration has so far consisted largely of voluntary agreements. Major AI labs have signed several pledges with the US government to proactively share concerns about models and work with external auditors, but those pledges are not binding, which effectively leaves the industry to self-regulate. Third-party testing is not required under law. The FTC subpoenaed OpenAI and Anthropic over consumer harm concerns, a step that followed a rapid shift in which the government said it would start requiring AI companies to disclose safety incidents and security breaches. That shift came after Anthropic revealed that Claude had reported a false homicide tip and filled out government forms. In a statement to Axios, the Super Intelligence Force said AI companies must immediately disclose incidents involving their models and follow with swift action to remedy harm. It remains unclear what about that particular incident prompted the requirement, given a summer of high-profile breaches by frontier lab models during testing that were comparatively more destructive.
Whether Anthropic is simply executing its own prior plan or trying to get ahead of the FRONTIER Act is difficult to determine without knowing what lobbying preceded both the announcement and the government's change in posture.
Why it matters: For teams building on or buying frontier models, the practical effect is a new, checkable source of information about model risk that does not come from the lab's own marketing. If the arrangement spreads, evaluation results could become a de facto input to procurement and compliance decisions even without a legal mandate. It also puts pressure on rival labs: matching this level of access would mean exposing internal systems to outsiders, while declining invites the question of what they are unwilling to show. For Anthropic, the upside is credibility; the risk, which is inference on my part, is that published findings may surface problems the company would rather have handled privately.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
As AI labs continue to self-regulate, external evaluators play a crucial role in spotting risks.