OpenAI publishes a policy blueprint for AI-enabled cyber defense

What is changing?

OpenAI published a 16-page policy blueprint on October 21 describing a limited “defender’s window” in which frontier AI could give cyber defenders an advantage. The document says AI can accelerate vulnerability discovery and response while also increasing the risks of misuse and agents acting beyond their intended boundaries.

The blueprint calls for clearer responsibilities across developers, inference providers, deployers, and users; monitoring for dangerous behavior throughout development and deployment; and reliable shutdown controls for operations each organization controls. It also covers secure-by-design practices for critical services, modernization of legacy systems, incident reporting, and standards for agent identity.

OpenAI also proposes public pre-deployment evaluations and independent technical assessments for frontier models and representative agent systems. The recommendations draw on lessons from incidents involving OpenAI’s own models, but they are a policy proposal for governments and industry rather than binding regulation.