Anthropic grants six METR researchers employee-level access — TechCrunch

Anthropic has granted six researchers from the nonprofit organization METR employee-level access to the company’s internal systems. They will be able to independently assess Anthropic’s models, its safety practices, and its approaches to risk management, TechCrunch reports.

METR tests advanced artificial intelligence models. The researchers received Anthropic-issued work laptops, office space, and access to internal tools roughly comparable to that available to the company’s risk assessment teams. Anthropic is also encouraging its employees to share information with the METR team in the same way they would with colleagues working on AI alignment risks.

Independent publication of assessments

METR will publish its own assessments of Anthropic’s risks, incidents, and practices. The company has waived its right to block such publications, but may request the removal of information only from narrowly defined categories. These include customer data, information protected by attorney-client privilege, and details about model architecture.

More current news is available on the UA.News Telegram channel Telegram.

At the same time, METR will be able to publicly explain the nature of such removals and express its position on their justification without disclosing the redacted information. Anthropic compared this model to banking oversight practices, in which regulators place permanent inspectors in the largest banks.

Three levels of external assessment

METR is so far the only evaluator at the highest level of the three-tier system Anthropic is creating. At the lower level, the biosafety nonprofit SecureBio, among others, conducts narrow reviews and requires Anthropic’s permission to publish the results. Mid-level evaluators receive more information and may publish results without the company’s approval. The highest level provides METR with access close to that of Anthropic employees.

The announcement implements an idea proposed by Anthropic CEO Dario Amodei in a September essay: bringing external evaluators directly into AI laboratories. OpenAI has also promised to introduce a similar approach. Anthropic, OpenAI, Google, Meta, Nvidia, and xAI signed a White House agreement to conduct independent evaluations.