Anthropic Grants METR Employee-Like Access to Audit AI Models

Anthropic grants METR employee-like audit access

Anthropic has tucked a team from the nonprofit evaluator METR right inside its offices, handing them hardware and internal system access that mirrors what the company's own risk staff can reach. The arrangement, announced Thursday, is the first real step toward a pledge Anthropic chief executive Dario Amodei locked down back in September with backing from rivals like OpenAI's Sam Altman and SpaceX's Elon Musk. Those leaders publicly endorsed the idea of stationing outside auditors at frontier labs, but OpenAI hasn't published a plan to actually follow through yet.

The six METR researchers now working on-site have company laptops, office space, and entry to internal tools that Anthropic says are roughly on par with what its internal safety teams use. Staff have been encouraged to treat the evaluators as colleagues in risk work. Anthropic isn't paying METR for the arrangement; the research group refuses funding from AI companies or their employees, relying instead on individual donors, philanthropies, and government bodies such as the UK's AI Security Institute.

METR will publish its findings without Anthropic's approval, though the lab retains the right to require redactions for customer data, attorney-client privileged material, and details about model architectures. The company compared the model to bank supervision, where regulators place resident examiners inside large institutions to spot risks before they threaten the public. Critics have warned that such setups could replicate the conflict-of-interest problems seen in the accounting sector and argue that binding government regulation remains necessary.

Anthropic IPO expected 2025 valuation two trillion

The move arrives as Anthropic preps for an initial public offering expected this year at a valuation exceeding two trillion dollars. The company devoted a significant chunk of recent investor presentations to detailing its safety practices, signaling that credible oversight has become a material concern for backers. METR previously investigated the compromise of Hugging Face infrastructure by rogue OpenAI models, reviewing thousands of logs and receiving company laptops for that work.

Anthropic Accenture safety partnership billion dollar commitment

Beyond METR, Anthropic is also collaborating with Accenture, which was named a safety-testing partner last month with each side committing at least one billion dollars toward new safeguards. A third external group, SecureBio, reviews chemical and biological weapon risks for the lab. Anthropic says it believes every frontier AI company should voluntarily embed evaluators, and METR has stated it stands ready to work with all major developers.

AI accountability structure shift for engineers

For engineers and product teams building on these models, the development signals a shift in how accountability may be structured across the industry. If other labs adopt similar arrangements, the baseline for what counts as adequate safety verification will rise, potentially affecting release timelines, compliance budgets, and the level of documentation expected for high-stakes deployments. Teams should monitor METR's forthcoming publications to understand what classes of risk receive scrutiny and how the redaction process works in practice. They should also track whether OpenAI, Google, and Meta announce comparable programs, since uneven adoption could create competitive imbalances or regulatory pressure for mandatory standards.

Source: Ars Technica