Critical cybersecurity capability is a threshold used in AI safety to describe a model that may be able to carry out advanced cyber offense tasks with minimal human help. At this level, a system could potentially discover and adapt exploits, chain steps into a complete attack, or produce working zero-day exploit paths against hardened targets. The label is not about general coding skill; it is about offensive security power.
This matters because a model near this threshold changes the risk profile of development and deployment. Security teams may need stronger evaluation, tighter access to tools and networks, and slower release decisions. In attacks, such capability could automate reconnaissance, exploit development, and post-exploitation planning. In defenses, the same models can help analyze logs, triage alerts, and speed incident response, but only if they are governed carefully. The practical goal is to keep advanced AI from becoming an easy force multiplier for attackers before controls are ready.