AI Offensive Cyber Capabilities Are Doubling Every Six Months, Safety Researchers Warn
A new safety research report finds that AI models' ability to autonomously exploit security vulnerabilities has been doubling roughly every 5.7 months since 2024 — a rate that is outpacing the development of defensive tooling and policy frameworks.
D.O.T.S AI Newsroom
AI News Desk
The AI safety research community has been tracking the growth of AI-assisted offensive cyber capabilities for two years. The latest findings are stark: the ability of frontier AI models to identify and exploit real-world security vulnerabilities has been doubling approximately every 5.7 months since 2024. If that trajectory continues, the capability gap between AI-assisted attack and AI-assisted defense will widen significantly before any regulatory framework has the tools to close it.
What "Doubling" Means in Practice
The researchers evaluated AI models against a standardized set of known vulnerability classes — including SQL injection, privilege escalation, and memory corruption patterns — and measured the rate at which models could autonomously identify, adapt to, and exploit novel instances of each class without human guidance. In early 2024, frontier models could autonomously complete roughly 15% of these tasks. In early 2026, the figure sits above 50% and is still rising.
The 5.7-month doubling rate is not a smooth curve. Safety researchers note it reflects a series of step-changes tied to model capability releases, particularly the introduction of extended context windows and longer chain-of-thought reasoning, which allow models to trace multi-step vulnerability chains through large codebases that would have previously required expert human analysis.
The Defense Gap
Cybersecurity vendors are actively building AI-assisted defensive tooling, but the research suggests offense is currently moving faster. Detection systems are trained on historical attack patterns; AI-generated attacks can be novel enough to evade signature-based defenses while still being systematic enough to be highly effective. The practical implication is that organizations relying on traditional security monitoring may face an increasingly asymmetric threat environment.
For AI labs, the finding adds urgency to the debate over how capability evaluations should handle dual-use risks. Anthropic, OpenAI, and Google DeepMind all conduct pre-release safety evaluations that include cybersecurity assessments. The research suggests those assessments may need to be updated on shorter cycles than current release cadences allow.
