Anthropic is letting METR researchers look under the hood with employee-level access

Anthropic is giving employee-level entree to six researchers from METR, a nonprofit that tests frontier AI models, to independently measure Anthropic’s models, information practices, and hazard management, the institution said successful a blog post connected Thursday.

The statement comes astatine a clip wherever incidents of AI agents acting extracurricular their instructions person been disclosed every week or two since precocious August. The incidents person sharpened calls for autarkic checks connected AI companies.

“We privation to beryllium publically accountable for the decisions we’re making,” Anthropic said successful the blog post. “We purpose to adjacent the spread betwixt what evaluators recognize astir a frontier AI company’s systems and what the institution itself understands.”

Under the caller arrangement, the METR squad has Anthropic-issued institution laptops, bureau space, and entree to interior tools astir connected par with what Anthropic’s ain hazard appraisal teams use. “Anthropic employees are encouraged to stock accusation with them conscionable arsenic they would with a chap worker moving connected alignment risk” the institution said.

Anthropic compared the statement to banking, akin to however regulators spot nonmigratory examiners wrong the largest banks. 

METR volition people its ain assessments of Anthropic’s risks, incidents, and practices, and the institution has fixed up the close to artifact them.

“Anthropic can’t veto publication, but we tin necessitate them to redact accusation if it falls into narrow, pre-defined categories,” the institution said, specified arsenic lawsuit data, thing attorney-client privileged, oregon relating to accusation astir its exemplary architectures. “METR is escaped to qualify our redactions publicly, and accidental whether it believes the redactions were justified, arsenic agelong arsenic it does truthful without revealing the redacted information” Anthropic added.

METR is the lone evaluator truthful acold astatine the apical of a three-tier strategy Anthropic is building. Evaluators successful the bottommost tier, similar biosecurity nonprofit SecureBio, tally constrictive reviews and necessitate Anthropic’s support to publish. The mediate tier gets much accusation and tin people without approval, portion the apical tier adds METR’s employee-level access. 

Thursday’s announcement makes bully connected a committedness Anthropic CEO Dario Amodei made in a September essay, wherever helium projected embedding extracurricular evaluators wrong AI labs arsenic a archetypal measurement toward slowing the technology’s development. OpenAI rapidly pledged the same, and some companies, on with Google, Meta, Nvidia and xAI, person since signed a White House accord committing to autarkic review.