Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman have proposed embedding third-party safety evaluators directly into their frontier AI development processes. This initiative aims to grant external researchers, such as METR and Redwood Research, deeper access to internal systems, training checkpoints, and logs to assess model alignment and report safety incidents.
The proposal suggests that evaluators should have the authority to publish findings on risk levels and internal practices without editorial oversight from the AI companies. Researchers note that current testing methods are often limited by short timeframes, such as the three-day window provided for testing the GPT-6 Astra model, which complicates the ability to draw definitive conclusions about safety.
Industry experts emphasize that the effectiveness of this model depends on whether companies are willing to relinquish control over the evaluation process. Historically, third-party firms have faced restrictive non-disclosure agreements and limited access, often being treated as standard contractors rather than independent watchdogs.
While Meta, SpaceXAI, and Google DeepMind have not committed to this specific embedding approach, regulatory frameworks are evolving. California’s SB 813 now establishes a system for state-recognized independent verification organizations, and the EU AI Act mandates that frontier developers conduct and document adversarial testing.
