Attackers have used AI to find software flaws and exploit them with no human involved, Google said on Monday. It says this is the first time it has seen it. The finding is in the latest AI threat report from the Google Threat Intelligence Group (GTIG) and Mandiant.
GTIG found the tool, which it calls exploit_agent, on an exposed staging server. It works with several frontier and locally run AI models. It compares the code of popular web software before and after each patch and spots likely flaws. Then it writes exploits and runs them on its own.
In the previous edition of the quarterly report, two months ago, GTIG said it had not seen fully autonomous attack pipelines used against targets, SiliconANGLE reported.
The server also held data stolen from victims. It had scripts for brute-forcing passwords, installing webshells and removing rival attackers’ webshells. A configuration file carried comments in Russian. Victim data suggests the operator was targeting French media, entertainment and education websites, GTIG said.
Patched flaws may be easier targets for AI than unknown ones, Google said. Attackers can study a finite set of material: old and new code, release notes and public research.
It pointed to two flaws in PaperCut print servers that attackers chained together in late August. Researchers and the security firm Huntress used AI to build working exploits within a day of disclosure. Four days after disclosure, an attacker likely used AI to run a campaign against hundreds of organisations, according to GreyNoise.
After the PHP flaws were patched in October, hundreds of proof-of-concept exploits appeared that did not work. AI likely made many of them, the report says.
Several groups now use open-source AI hacking frameworks in live attacks, GTIG said. They include financially motivated gangs and suspected Chinese spy groups. It saw the CAIRN framework used against targets including the US aerospace sector.
These tools are noisy and leave traces that make them easier to spot, Google said. It expects attackers to rely less on humans and to become stealthier as the tools improve.
“We believe it is highly likely that threat actors will continue to use these frameworks in campaigns targeting organizations globally,” GTIG wrote.
The report also describes attackers who steal cloud and AI credentials to run models on victims’ accounts. In some cases, the fraudulent charges topped $100,000 in a single day, it says. One attack used a flaw in the LiteLLM gateway . A suspected North Korean group changed Claude Code settings files inside companies so that each launch downloaded a backdoor. Google says it disabled the accounts and assets linked to misuse of its Gemini models.
Get the most important tech news in your inbox each week.