Anthropic wants outside AI auditors to get employee badges and laptops

Anthropic CEO Dario Amodei proposed giving independent AI evaluators permanent, employee-level access to frontier laboratories, including office desks, access badges and company laptops. Anthropic says it will adopt the arrangement itself as the first step in a broader plan to slow capability gains enough for safety work to keep up.

Amodei published the proposal on September 12 in an essay titled We Must Pace the Frontier. He argued that AI development is entering a phase where models help build their successors, creating the possibility that progress accelerates faster than researchers can understand or control it.

An auditor who works inside the building

External model tests are usually conducted around releases or through limited access programs. Amodei's model would place third-party evaluators inside a company on an ongoing basis.

The evaluators would receive access similar to Anthropic's internal risk staff. They could inspect whether safety measures are being followed, report incidents and assess a model's alignment while it is being trained, not only after a finished system is presented for testing.

The unusual part is the proposed independence. Amodei said evaluators should be able to publish their findings without company editorial control. Anthropic could request narrow redactions for security or legal reasons, but the evaluator could disclose that a redaction had occurred and whether it believed important information was being hidden.

He pointed to METR, a nonprofit that evaluates advanced AI systems, as an example of the kind of organization that could fill the role. The final structure and the identity of the first embedded team have not been announced.

Slowing down without stopping

Amodei is not calling for a permanent halt to AI training. He describes "pacing" as deliberately moderating the speed of capability advances so safeguards, evaluations and public institutions have time to respond.

His proposal has three layers. Individual laboratories would first accept deeper outside scrutiny. Companies in democratic countries would then coordinate around common safety expectations. The final and most difficult step would be international coordination, including countries whose firms might otherwise gain an advantage by moving faster.

That structure reflects the central collective-action problem: a company may believe slower development is safer while fearing that a rival will simply use the extra time to pull ahead.

Rivals are already responding

OpenAI CEO Sam Altman said his company would also commit to independent evaluators with employee-like access and promised more details. Hugging Face CEO Clément Delangue said his organization had asked to participate in Anthropic's embedded-evaluator program.

Those responses are statements of intent, not operating audit programs yet. The test will be whether the evaluators receive meaningful access, whether their reports appear without company filtering and whether competing laboratories adopt comparable rules.