Anthropic CEO Dario Amodei published an essay September 12 proposing that frontier AI labs deliberately slow capability gains and embed third-party safety evaluators inside their companies. The evaluators would get access comparable to an internal risk team: desks, badges and company laptops.
OpenAI CEO Sam Altman said OpenAI would do the same, and xAI's Elon Musk responded "Dario is right," signaling agreement despite xAI's positioning as the industry's least safety-constrained lab, according to Technori's account of the exchange.
The proposal builds on AEF-1, a standard published by the AI Evaluator Forum and last updated December 4, 2025, that sets five minimum operating conditions for independent evaluations: sufficient access and resources, minimized conflicts of interest, analytic autonomy, transparent methods and results, and protection of sensitive information. Evaluators including Transluce, METR and SecureBio already publish compliance checklists against it, according to the forum's own site.
Evaluators told TechCrunch the access has been the real problem. Apollo Research said it received only three days to assess an OpenAI model referred to as GPT-6 Astra, which "did not provide substantial evidence about the model's alignment." FAR.AI's Adam Gleave said his firm has "had to turn down contracts with several frontier developers that wanted too much control over the evaluation process," and Safer AI's Henry Papadatos said voluntary measures "are always dependent on a company's goodwill," calling for legislation instead.
Meta and Google DeepMind have not committed to the practice; DeepMind CEO Demis Hassabis has proposed a separate industry standards body instead, TechCrunch reported.
For any company building on these models, an evaluator's findings are only as good as the access and editorial control behind them, the same test any outside audit has to pass before its conclusions mean anything.