AI has already helped hackers make cyberattacks faster and easier to scale. Now, things are going a step further.

Google says it has identified the first instance of threat actors fully automating vulnerability discovery and exploitation using AI, without a human in the loop.

The discovery marks a significant shift in how AI agents could be used to develop and execute cyberattacks.

IBM AI Cybersecurity Report 2026: AI-Driven Attacks Up 56%
A new data breach report shows AI is driving more sophisticated cyberattacks while helping organisations detect breaches faster and save nearly $2 million per incident.

How Google's exploit_agent framework works

Google's Threat Intelligence Group and Mandiant found a multi-agent framework called exploit_agent on an exposed staging server. It connects multiple AI agents to frontier and locally hosted language models.

The system continuously compares widely used software versions before and after vendors release security patches. By analysing code changes, it can identify potential vulnerabilities, draft exploits, and execute them automatically.

Investigators also found Russian-language comments in its configuration file, including instructions for setting up LLM API keys for automated exploitation. The comments referred to it as an “exploit_agent pilot”. However, Google did not attribute the framework to a Russian threat actor or any specific group.

Data recovered from the server suggested that its operator was interested in French websites in the media, entertainment, and education sectors. Investigators also found scripts for reconnaissance, password guessing, and installing webshells, including one designed to remove those belonging to rival attackers.

Although Google found data exfiltrated from targeted victims on the server, it did not explicitly establish that exploits generated by exploit_agent were responsible for those breaches.

Other AI frameworks are already appearing in cyberattacks

Google also identified threat actors using open-source offensive frameworks such as ARTEX, CAIRN, CyberStrikeAI and PentAGI.

It independently observed CAIRN and PentAGI deployed in real-world activity. CAIRN was used against targets including a US aerospace organisation, while a customised version of PentAGI was used for reconnaissance and vulnerability scanning against organisations across Asia.

Google also observed experimentation with CyberStrikeAI and HexStrike AI. Separately, a China-linked group was developing its own automated attack framework, likely for espionage.

How AI is shrinking the window between patches and exploits

Subscribe for free to continue reading this article

Subscribe Subscribe

Already have an account? Log in