Strengthening cyber resilience as AI capabilities advance

OpenAI says its models are rapidly improving on cyber benchmarks, including capture-the-flag and exploit-related tasks.

The company expects future frontier models to reach high cybersecurity capability levels and is preparing stronger safeguards.

Planned protections include defense-in-depth controls, monitoring, refusal behavior, and enforcement to curb misuse.

The goal is to preserve legitimate defensive use while reducing the risk that advanced AI helps attackers.