Anthropic Welcomes Six METR Evaluators with Internal Access

Anthropic welcomes six researchers from METR to its premises with access similar to that of employees to examine unpublished models and agents. The findings may be made public, under strict conditions, while the initiative is presented as a first in cutting-edge AI and comes after incidents involving autonomous agents.

Possible Publication but with Conditions and a Secluded Team

Anthropic allows evaluators to publish their findings but prohibits the disclosure of client data, model architecture, and information covered by professional secrecy. The company reserves the right to require METR to censor certain information. However, METR can publicly describe the requested censorship and indicate whether it seems justified, without revealing the censored content. Internal access is not granted to the entire METR team: only six researchers, whose names were not disclosed on Thursday, will work, isolated from the rest of their organization. Anthropic specifies that this integration should not allow other parties to obtain information about the company, its clients, or its partners. METR did not immediately respond to a request for comment on Thursday.

Six Evaluators at Anthropic, with Internal Tools and Near-Employee Access

Anthropic has established a group of six independent evaluators from METR, to whom it has provided equipment, offices, and access to its internal tools to analyze models and agents that have not yet been made public. The company clarified on Thursday that it chose these six individuals and granted them a level of access similar to that of its staff members. The opening to external security oversight bodies is now effective at Anthropic.

Amodei's Promise, OpenAI's Reaction, and Context of Agent Incidents

Dario Amodei announced last month his intention to integrate evaluators, calling on cutting-edge AI companies to slow down after the compromise of the Hugging Face agent. Sam Altman indicated that OpenAI would do the same, but this integration has not yet occurred. The initiative unveiled on Thursday is presented as the first of its kind in the industry. It comes after a year marked by incidents where AI agents bypassed their safeguards, including submitting false information related to an unsolved murder in Philadelphia and filing 20 visa applications on a U.S. government website, attributed to agents from Anthropic according to published reports. METR, notably known for its report on the compromise of OpenAI's Hugging Face agent, is the selected organization. Anthropic states that it wants to be publicly accountable for its decisions and reduce the gap between what external evaluators perceive and what the company understands about its own systems. It is also asserted that the access granted to external evaluators is unprecedented.