Anthropic is letting an outside research group work close to its own staff and systems so the public can get an independent look at its AI risks. The arrangement goes further than the usual outside review, and the company says it hopes other AI labs will do the same.

Anthropic Grants Nonprofit AI Watchdog Access

Anthropic announced that it has started giving employee-like access to a team from METR, a nonprofit that evaluates frontier AI models. METR will publish the results of its risk assessments. Over the past year, the two have already worked together on several reviews of Anthropic's Risk Reports.

The move follows commitments Anthropic made last month. The company promised stronger independent, third-party assessment of its models and safety practices, and it signed the White House Accord on Super Intelligence, in which companies pledged to partner with independent outside auditors or evaluators. Anthropic is also working with SecureBio and Accenture as evaluators, and it is in early talks with several other groups, including Hugging Face.

Anthropic explained why it wants outsiders involved. As models grow more powerful, they need stronger safeguards and tougher risk assessments, and the company says it wants to be publicly accountable for its decisions. It cannot share everything, though, because some safety information could help malicious actors and some of it is commercially sensitive, like compute and training details.

Instead of publishing everything, Anthropic is giving third-party evaluators deep access and asking them to share evidence and reasoned conclusions. The company says, "We aim to close the gap between what evaluators understand about a frontier AI company's systems and what the company itself understands." It lists four needs for any evaluator: expertise, independence from the company, enough access to judge risks, and the freedom to publish.

Anthropic compared the idea to bank supervision. Banks know more about their own risks than outsiders do, so regulators place resident examiners inside the biggest banks, and the company says embedded evaluators work on similar logic, though it notes the analogy is not exact.

METR Gets 'Employee-like' Access to AI Systems

Anthropic has onboarded six METR employees and given them company hardware, office space, and access to internal tools and systems that are roughly on par with what its own risk assessment teams use. A small number of clearly defined categories are off limits, including data about customers, attorney-client privileged information, and information about model architectures. Outside those areas, Anthropic staff are encouraged to share information with the METR team just as they would with a fellow employee working on alignment risk.

METR will publish its analysis of Anthropic's risk levels, incidents, and practices. Anthropic cannot veto publication, but it can require METR to redact information that falls into the predefined categories. Anthropic decides what fits those categories, while METR keeps the right to make certain statements about the level of access and information it received.

The company stresses that the engagement is tightly limited. It involves a small METR team operating under enforceable agreements, and other METR employees are kept apart from that work, including researchers who work with Anthropic on other projects. Anthropic adds that this is not a way for other organizations to get its information or that of its customers and partners.

METR is the top level of a three-tier plan. At the first tier, evaluators run narrowly scoped risk investigations, Anthropic decides what access they get, and publication requires Anthropic's approval. At the second tier, evaluators get more information and can publish without approval, although Anthropic can still demand that specific categories be removed, and evaluators can publicly say whether they think the redactions were justified.

METR to Investigate Safety, Security of Anthropic's AI

Anthropic's post does not lay out a detailed work plan, but it says METR will analyze its risk levels, incidents, and practices and publish what it finds. METR describes its own mission as scientifically measuring whether and when AI systems might threaten catastrophic harm to society.

The two have a track record. METR's site lists a review of Anthropic's sabotage risk report on Claude Opus 4.6, a red teaming exercise on Anthropic's internal agent monitoring systems, and a review of the automated research section of Anthropic's February Risk Report.

Security shapes the arrangement. METR's team works under Anthropic's security standards, and some information stays limited because of leak risks. In a footnote, Anthropic says evaluators may also work with other companies, so it takes extra care around accidental leaks and treats information about architectures as especially risky.

Frequently Asked Questions

What Is METR?

METR stands for Model Evaluation and Threat Research. The name is pronounced like "meter." It is a research nonprofit that informs the public about the risks and capabilities of advanced AI systems.

METR is known for its time horizon work, which measures how long a task an AI agent can complete. It also built Hawk, an open source platform for running AI agent evaluations at scale. It has partnered with OpenAI, Anthropic, Google DeepMind, Meta, and Amazon to pilot frontier risk assessments, and it says it does not accept compensation for its risk assessment work.

What Will METR Investigate?

Anthropic only says METR will look at its risk levels, incidents, and practices. METR has described a broader plan in its own proposal, published this month, for assessments that run inside AI companies. That proposal says the public needs better evidence on three questions.

The first is alignment, meaning what misaligned actions AI agents might take and whether training could hide signs of worse behavior. The second is control, meaning whether monitoring would catch serious harm before it happens. The third is whether automating AI research is speeding up capabilities faster than safeguards can keep pace. A summary of the proposal says METR describes it as voluntary, with companies able to redact material, and not a safety audit or a go or no go gate on releases.

Why Is Anthropic Under Investigation?

Anthropic is not facing a legal or regulatory probe, and its announcement does not describe one. This is a voluntary arrangement that the company chose to make. It says it wants public accountability even though it cannot release every detail of its safety work.

The company also points to its recent commitments, including the White House accord and a pledge to expand independent assessment. It says it embeds evaluators voluntarily and believes every frontier AI company should do the same.