Six external researchers from METR now carry official Anthropic laptops, occupy desk space in the company's offices, and hold credentials for internal diagnostic tools. Their brief is unusual for frontier tech: inspect Anthropic's systems from the inside, audit risk management, and publish whatever they discover without corporate sign-off.
Anthropic announced the arrangement on Thursday, comparing the resident monitors to regulatory examiners embedded on commercial bank trading floors. The company told staff to treat the outside team like colleagues working on alignment risk, encouraging internal teams to share technical details openly.
The lab surrendered any editorial veto over METR's output.
METR will be allowed to independently publish assessments detailing Anthropic's risks, internal incidents, and testing practices. Anthropic can only mandate redactions across narrow, pre-arranged boundaries — specifically customer records, proprietary model architectures, and communications shielded by attorney-client privilege. Even then, METR retains the right to state publicly whether it considers those corporate redactions justified.
Thursday's move translates an argument Anthropic CEO Dario Amodei published in September into actual company policy. In an essay examining safety bottlenecks, Amodei proposed putting independent evaluators directly inside frontier labs as a practical mechanism to decelerate irresponsible releases.
A Three-Tier System
The timing is deliberate. Autonomous software agents have slipped their guardrails in disclosed incidents every week or two since late August, intensifying calls for verifiable corporate transparency across the sector.
Anthropic is structuring its outside audits across three tiers, and METR currently sits alone at the top.
At the bottom tier, groups like biosecurity nonprofit SecureBio perform narrow assessments and require Anthropic's explicit sign-off before publishing anything. A middle tier grants wider technical access and allows public reporting without corporate clearance. The top tier gives METR full employee-level visibility into internal systems.
Pressure now shifts toward the rest of the industry. OpenAI pledged a comparable arrangement immediately after Amodei published his September essay. Both companies, along with Meta, Google, Nvidia, and xAI, have since committed to independent outside reviews under a White House accord. Whether those competing labs will match Anthropic by handing outside auditors actual company hardware and internal accounts remains the next question.
ADFiled by The AI Desk
Models, assistants and the companies and chips behind them, reported from what was released and what was claimed, with the difference kept clear.
More from this desk →
Be the first to comment
Join the argument. No password, just your email or a passkey.