Strengthening Cyber Resilience as AI Capabilities Advance

AI models' cybersecurity capabilities are advancing rapidly. Looking at CTF (Capture-the-Flag) challenge performance, GPT-5 scored 27% in August 2025, while GPT-5.1-Codex-Max showed significant improvement, reaching 76% in November 2025.

OpenAI is preparing for the possibility that future models will reach 'High' level capability, where they could develop zero-day remote exploits or support complex enterprise intrusions. In response, OpenAI is focusing investment on developing tools that make it easier for defenders to perform code audits and patch vulnerabilities, as well as on defensive cybersecurity work.

Given the dual-use nature of offensive and defensive techniques, which are similar, OpenAI is implementing the following multi-layered security (defense-in-depth) strategy:

• Model training: Training models to be useful for educational and defensive purposes while refusing clearly malicious cyber abuse requests.
• Detection systems: Detecting malicious activity through monitoring across products, blocking outputs or routing to safer models when risk is identified.
• Red team operations: Collaborating with specialized organizations to continuously evaluate and improve security mitigations.