Google Sees AI Agent Develop Exploits Without Human Direction
Google Threat Intelligence Group says it has observed an AI system move beyond assisting an operator and begin developing software exploits on its own, a threshold that could compress the time between a disclosed vulnerability and active exploitation.
In its latest AI Threat Tracker, Google described a framework called exploit_agent that continuously compared code before and after a security patch, identified the underlying vulnerability, and produced exploits without human intervention. The company linked the activity to tooling used by Russian threat actors.
The finding does not mean AI has replaced skilled attackers. It does show that autonomous systems can now handle a technically demanding portion of the attack chain that once required repeated human analysis. For defenders, that shifts the risk calculation around patch disclosure, proof-of-concept code, and internet-facing systems.
Google also reported seeing open-source agent frameworks such as CAIRN, PentAGI, and CyberStrikeAI used in live intrusions. Other observed uses included influence operations, North Korean operators using AI-generated personas, and attempts to alter model configuration.
The practical takeaway is speed. Security teams may have less time to test and deploy patches before automated systems can translate a code change into an exploit path. Vulnerability management programs should prioritize exposed assets, reduce approval delays for emergency fixes, and monitor for reconnaissance that follows major vendor patches.
Google's report is based on the company's threat intelligence visibility and should be read as an assessment from the vendor. Even so, the autonomous exploit-development example is a concrete sign that agentic offensive tooling is becoming operational rather than merely experimental.


