Study Finds AI Offensive Cyber Capabilities Doubling Every Six Months

New research reveals that AI models' offensive cyber capabilities have been doubling every 5.7 months since 2024, marking a rapid acceleration in automated security threat potential. Advanced models including Claude Opus 4.6, GPT-4o, and GPT-5.3 Codex can now solve expert-level cybersecurity tasks that typically require human experts approximately three hours to complete. The findings demonstrate AI systems are rapidly improving at exploiting security vulnerabilities, raising concerns about the growing gap between offensive AI capabilities and defensive measures. This exponential growth trajectory poses significant challenges for cybersecurity professionals and organizations attempting to protect systems against increasingly sophisticated AI-powered attacks.

Rapid acceleration of AI offensive cyber capabilities

• ▪AI models are rapidly improving at exploiting security vulnerabilities.
• ▪AI models' offensive cyber capability has been doubling every 5.7 months since 2024.

Current AI model performance in cybersecurity tasks

• ▪Opus 4.6 can now solve cybersecurity tasks that take human experts about three hours.

Perspective of AI safety researchers

• ▪AI models achieving expert-level cybersecurity task performance in three hours poses risks for automated exploitation of vulnerabilities.
• ▪The 5.7-month doubling time for AI offensive cyber capabilities represents an unprecedented acceleration in security threats.

Perspective of Cybersecurity industry

• ▪Traditional cybersecurity defenses may become inadequate against AI models that can exploit vulnerabilities at expert-level speed.
• ▪The cybersecurity industry must accelerate development of AI-powered defensive tools to counter offensive AI capabilities.

1 source