Google identifies autonomous AI cyberattack
Google has identified what it described as the first observed cyberattack in which an AI-powered system automated vulnerability discovery and exploitation without human intervention. An exposed staging server hosted a multi-agent framework called exploit_agent, connected to advanced and locally hosted language models. It could examine software updates, compare code around security patches, identify weaknesses, develop exploits and execute them. Supporting scripts enabled internal network reconnaissance, brute-force attempts and the installation of webshells to maintain access. The operator appeared interested in French media, entertainment and education websites.
Threat actors are increasingly adopting open-source frameworks that assign different intrusion tasks to cooperating AI agents, Google said. These systems can generate attack code and adjust their actions using feedback from previous attempts. Automating comparisons between patched and older software also helps attackers develop exploits against organisations that have not installed updates, shortening the time available to secure vulnerable systems.
Google also observed AI supporting espionage, social engineering and influence operations, including synthetic personas and automated political messaging. Separately, criminals are stealing cloud credentials to access and resell AI resources, creating substantial fraudulent charges. Despite their growing capabilities, autonomous attack tools remain imperfect: their activity can leave identifiable traces, and generated exploits do not always work as intended.