Attackers used an artificial intelligence system to find weaknesses in newly patched software, write working exploits and run them without human involvement, Google said Monday, describing the activity as the first fully autonomous attack pipeline it has observed.
The finding came from the Google Threat Intelligence Group and Mandiant, which identified a tool called exploit_agent on an exposed staging server. Google said the system can work with several frontier and locally run AI models and is designed to compare software before and after a patch, identify likely flaws and then create and execute an exploit.
How the system operated
The tool focused on software changes made through security updates. By examining the code before and after a patch, along with release notes and publicly available research, it could narrow its search to a defined set of changes. Google said patched vulnerabilities can be easier for AI systems to study than previously unknown flaws because the relevant technical material is available for comparison.
The server also contained data taken from victims and scripts for guessing passwords, installing web shells and removing web shells belonging to other attackers. A configuration file included comments in Russian. Google said information from the victims indicated that the operator was targeting websites in the French media, entertainment and education sectors.
- Exploit_agent was found on an exposed staging server.
- The tool compared software versions before and after patches to locate likely vulnerabilities.
- Google said the system generated and ran exploits without a person directing each step.
Recent examples involving patched flaws
Google pointed to two vulnerabilities in PaperCut print servers that attackers chained together in late August. Researchers and the cybersecurity company Huntress used AI to produce working exploits within a day of the flaws being disclosed, Google said. Four days after disclosure, GreyNoise assessed that an attacker likely used AI in a campaign against hundreds of organizations.
Google also said that hundreds of proof-of-concept exploits appeared after PHP flaws were patched in October. Many of those examples did not work, and the company said AI likely produced a substantial number of them. The activity illustrates how quickly attackers can test public technical information after a software maker releases a fix.
AI tools already used in attacks
Google said several groups are using open-source AI hacking frameworks in live operations, including financially motivated criminal groups and suspected Chinese espionage groups. Its researchers observed the CAIRN framework being used against targets that included the U.S. aerospace sector.
Google said the tools currently generate enough noise and leave enough traces to make them easier to detect. The company expects attackers to use fewer human operators and seek more concealed methods as the systems improve. The disclosure adds an autonomous exploit-generation tool to the cybersecurity activity Google is tracking in the wild.