OpenAI recently completed a three-week pilot program granting independent evaluators unprecedented access to its internal infrastructure to expand external oversight of advanced artificial intelligence (AI), non-profit safety groups METR and Redwood Research announced Thursday.
The initiative, which ran from July 21 to Aug. 11, allowed one researcher from each organization to inspect OpenAI’s internal codebase, documentation, and technical personnel. Operating on-site with chaperones and company-issued hardware that remained on the premises, the evaluators conducted a rigorous stress test of OpenAI’s internal misalignment monitoring system — a critical framework designed to ensure AI behavior remains aligned with human values and intent.
During the exercise, researchers uncovered several “novel vulnerabilities,” including an exploit that could permit automated agents to bypass monitoring for hundreds of recent actions. OpenAI confirmed it has since patched the identified security flaws affecting its production safeguards.
Due to the pilot’s scope, METR and Redwood released only a high-level public summary, while delivering a comprehensive 61-page report directly to OpenAI leadership. The organizations indicated they intend to share broader details regarding future collaborative assessments.
“This collaboration builds on our longstanding commitment to pioneering new approaches to third-party testing, and we’re using what we learned to strengthen our safeguards and shape future collaborations with independent evaluators,” an OpenAI spokesperson told CNBC.
The pilot marks a concrete step toward fulfilling a pledge made in September by OpenAI CEO Sam Altman to grant external auditors “employee-like access.”
The concept was initially proposed by Anthropic CEO Dario Amodei as a mechanism to promote safety and manage the acceleration of AI development. Anthropic has established similar partnerships with organizations such as METR, Accenture, and SecureBio, while OpenAI confirmed it is actively finalizing additional third-party assessment contracts.
External evaluation has gained momentum across both industry and government channels.
President Donald Trump has advocated for independent auditing, encouraging major AI firms to join a voluntary White House accord focused on safety standards. Industry leaders from Alphabet Inc.’s Google, Meta Platforms Inc., OpenAI, SpaceX, NVIDIA Corp., and Anthropic have signed onto the initiative.
Despite growing political and corporate backing, experts note that critical operational questions persist regarding how independent evaluator groups will maintain long-term funding and structured reporting independence from the tech giants they audit.
Simultaneously, OpenAI announced updates to its developer Usage Policies and safety frameworks as it deploys advanced systems like GPT-6 Astra. The company emphasized a multi-layered approach to safety, combining pre-deployment risk evaluations, adversarial red-teaming, and post-deployment abuse monitoring.
By integrating internal monitoring with enhanced external access programs, OpenAI aims to establish clearer baseline requirements for developers while maintaining consistent safety enforcement as model capabilities continue to expand.
