Over 100 AI experts call for independent safety evaluators at frontier labs

On September 18, 2026, a coalition of over 100 AI experts, including Geoffrey Hinton, published a letter demanding independent safety evaluations at frontier AI labs like OpenAI and Anthropic. Organized by the AI Evaluator Forum, the group calls for "employee-like access" to internal systems alongside robust protections, such as limited non-disclosure agreements and shields against corporate retaliation. While tech leaders like Sam Altman and Elon Musk support the concept, national security experts warn that relying on labs' self-assessments risks critical infrastructure.

AI evaluator coalition letter

• ▪Signatories of the public letter demanding independent safety evaluations at frontier AI labs, published September 18, 2026 by the AI Evaluator Forum, include AI pioneer Geoffrey Hinton, members of Johns Hopkins University and Stanford University, and the nonprofit evaluator METR
• ▪A public letter demanding independent safety evaluations at frontier AI labs, published September 18, 2026, was organized by the AI Evaluator Forum, a consortium chaired by Conrad Stosz, and shared exclusively with CNBC
• ▪On September 18, 2026, a coalition of over 100 artificial intelligence experts, academics, and evaluators published a public letter calling for independent third-party safety evaluations of frontier AI models

Embedded evaluator access proposal

• ▪Conrad Stosz stated that Anthropic CEO Dario Amodei's proposal to give third-party evaluators 'employee-like access' could involve giving them access to company computers, sensitive internal data, unreleased systems, and candid employee communications
• ▪Anthropic Chief Executive Officer Dario Amodei proposed providing third-party safety evaluators with "employee-like access" to inspect and audit frontier AI models and development processes

National security concerns

• ▪Vinh Nguyen, a Council on Foreign Relations senior fellow, stated that independent evaluators are necessary because the public cannot depend solely on powerful labs' own accounts of safety
• ▪Vinh Nguyen warned that a few powerful labs control capabilities that can endanger national security, critical infrastructure, cybersecurity, and the broader economy

Debatable claims

• ▪Frontier AI labs should grant independent evaluators employee-like access
• ▪Embedded AI safety evaluators should be allowed to publicly release their findings