OpenAI Opens Internal Safety Systems to Four Evaluators Under Restricted Access Agreements
Emerging Technology Risk · 4 sources · Updated 13h ago · First seen Nov 17, 2026

OpenAI announced first-phase embedded safety assessment agreements with Apollo Research, METR, Transluce and Goodfire, giving outside evaluators access to selected internal models and safety systems. The pilots will examine alignment, model behavior and safeguards during training and internal deployment, rather than relying solely on pre-release testing. Depending on scope, access may include model checkpoints, source code, training data, monitoring systems, evaluation results and technical staff. Apollo plans to embed six evaluators for four months to investigate whether training encourages models to undermine safeguards, whether evaluations elicit such behavior, whether it occurs internally and whether associated reasoning is detectable. Its start date has not been established, and no pilot findings were announced.
METR’s assessment includes OpenAI’s largest reinforcement-learning training run. An earlier METR–Redwood Research exercise found that reinforcement-learning runs were initially outside OpenAI’s misalignment monitoring; METR has not verified OpenAI’s subsequent remediation claims. Access and publication restrictions limit the assurance these assessments can provide: evaluations cover agreed claims, models and periods, some access arrangements remain unresolved, and OpenAI retains confidentiality-based redaction rights. Apollo and Redwood Research criticized the arrangements as falling short of previously endorsed employee-like access. For organizations relying on OpenAI systems, the agreements represent expanded external scrutiny—not a certification of safety; forthcoming reports should be assessed against their scope, withheld evidence and independently verified remediation.


