Google has discovered a tool designed to turn software patches into exploits automatically, without a human directing each step.

Google Threat Intelligence Group, the company’s cybersecurity research unit, described the finding in its Nov. report, “GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold.” Researchers located the software on an exposed staging server.

Called exploit_agent, the system integrates multiple frontier and locally hosted large language models. It is configured to compare web application code before and after patches, identify potential vulnerabilities and then draft and execute exploits. Google said this was the first time its researchers had observed attackers automating vulnerability discovery and exploitation without a human in the loop.

The server also contained stolen victim data and scripts for exploring internal networks, guessing passwords and installing webshells that enable remote access to compromised systems. Server telemetry suggested the operator was interested in targeting French media, entertainment and education websites. Google did not link the thefts to exploit_agent.

Configuration files contained Russian-language comments, including one describing exploit_agent as a pilot. Google made no attribution to a named group or the Russian government. It gave no victim count or evidence that the tool had conducted an end-to-end intrusion autonomously.

Earlier agents in live intrusions

In an Oct. 30 investigation, Google and Mandiant, its incident-response business, documented attackers using the ARTEX AI offensive tool in real intrusions. In several cases, operators supplied compromised credentials and instructed ARTEX to conduct subsequent attack stages.

In one operation, Google observed 27 events in 23 seconds of active burst time while the tool probed Amazon Web Services (AWS) permissions across security groups.

Unlike the previously documented intrusion activity, the newly discovered tool was configured to identify vulnerabilities by comparing patched and unpatched software. A screenshot in the November report shows exploit_agent inspecting access controls in the Spectra WordPress plugin. It does not establish that the plugin had a vulnerability or that an exploit succeeded.

Defending against automated attacks

Google’s findings raise a concern for organizations running internet-facing applications: attackers could use automated tools to identify vulnerabilities revealed by security patches before those updates are widely installed. The report does not measure how quickly exploit_agent could produce working exploits or establish whether its use shortened the time available for organizations to patch vulnerable systems.

In the October investigation, Mandiant recommended scanning exposed applications, separating internet-facing services from sensitive internal systems, restricting privileges and monitoring unusual bursts of administrative activity.

The November report also found that some autonomous offensive frameworks leave recognizable identifiers and code traces that increase their likelihood of detection. The report did not provide comparable detection findings for exploit_agent.

Personalized Feed
Personalized Feed