Google says attackers have automated an especially unwelcome form of code review: studying security patches to work out how to attack software that hasn’t been updated. In a November 16 report, Google Threat Intelligence Group (GTIG) described exploit_agent, an attacker-run AI framework that compares code, drafts exploits and executes them without a human intervening in that workflow. Source
GTIG says it first identified attackers fully automating vulnerability discovery and exploitation this way during the past quarter. Its earlier reporting described partially automated scanning and exploitation. Source
Researchers found the framework on an exposed staging server. It connects to multiple leading language models and locally hosted ones, though Google doesn’t name them. The server also held stolen victim data and scripts capable of surveying compromised networks, brute-forcing access, and installing webshells—tools for remotely controlling a server. Other scripts could remove rival attackers’ webshells. The toolkit even included a way to evict the competition. Source
Victim telemetry on the server suggested an interest in French websites in media, entertainment and education, according to GTIG. Google hasn’t identified the victims or attributed the framework to a named group. SourceSource
The fix doubles as a clue
The defensive headache is the gap between a fix becoming public and organizations installing it. GTIG says automated analysis can shorten the interval between disclosure and a usable exploit, leaving defenders less time to patch. A public fix can help attackers understand what was wrong with the older code. Source
Google argues that these disclosed, or “n-day,” vulnerabilities can offer language models a more manageable research task than hunting for previously unknown zero-days. Code changes, release notes and related research narrow the search. The patch gives defenders a repair—and attackers a pointed reading assignment. Source
Google hasn’t published a measured patch-to-exploit time, success rate or victim total for exploit_agent. Its account also doesn’t connect a particular generated exploit to a specific victim. Nor does it say whether the post-compromise scripts ran autonomously. SourceSource
Autopilot isn’t the whole operation
GTIG and Mandiant also observed attackers using other AI frameworks to automate particular phases and tasks during live intrusions. They describe the wider trend as approaching fully agentic intrusions. Source
Google says these broader frameworks “currently trade stealth for velocity”: unrestricted autonomous execution can generate conspicuous traffic, recognizable identifiers and distinctive code traces. It also warns that language models can invent details or even targets, reducing reliability. Source
These traces can help defenders detect the activity, but GTIG warns that the tools can still achieve their objectives. Source
Comments