OpenAI will allow independent groups to examine its AI models during earlier stages of development, the company announced on September 22. The shift is designed to surface safety problems before deployment, rather than after models are substantially complete.

Under the expanded evaluation program, independent reviewers will focus on four priority areas. First, they will assess OpenAI's "safety cases"—the company's own arguments for why a given model is safe to deploy. Second, evaluators will probe the resilience of critical safeguards against adversarial threats. Third, assessments will be tied directly to OpenAI's Preparedness Framework, the internal system the company uses to gauge catastrophic risks before launch. Fourth, evaluators will investigate misalignment incidents, a category that gained urgency after a breach involving Hugging Face highlighted how quickly such failures can escalate.

Because safety cases and the Preparedness Framework are both internally produced, opening them to outside reviewers adds an external check on how OpenAI itself judges risk before launch.

OpenAI also published seven guiding principles governing how these assessments should be conducted. The principles emphasize scientific rigor, evaluator independence, security protocols, and responsible methods for publishing findings. The company has been in discussions with organizations including METR and Redwood Research, both of which have previously collaborated with OpenAI on safety work. New partners have not been formally announced, and specific access terms remain under negotiation; how those terms are settled will help determine how much access evaluators actually receive.

The initiative builds on a commitment made by CEO Sam Altman on September 12, when he indicated that evaluators would be embedded more deeply within OpenAI's operational structure.

How OpenAI's Safety Approach Has Evolved

OpenAI's safety protocols have expanded considerably since the GPT-4 era, when the company first began inviting external red-teamers to probe its models before release. Those earlier efforts were narrower in scope, typically focused on the final stages before launch. The new framework moves that engagement significantly upstream in the development timeline, giving evaluators the opportunity to identify risks while there is still time to address them.

Talent and Competitive Considerations

The AI safety research community is relatively small, and organizations such as METR and Redwood Research are among the most credible voices in the field. By deepening relationships with these groups, OpenAI gains both practical expertise and reputational capital. For the evaluators, the arrangement offers equally valuable access to frontier systems that would otherwise remain behind closed doors.

The program's credibility will rest partly on what happens when an independent assessment surfaces findings that conflict with OpenAI's commercial interests. The seven principles explicitly call for responsible publication methods, but the tension between transparency and competitive advantage remains. If OpenAI navigates that tension credibly, the program could become a template for industry-wide safety standards. The near-term indicators to watch are concrete: which organizations are formally named as partners, what access terms are ultimately agreed, and how early findings are disclosed.