Google says attackers have automated an ugly bit of security homework: reading a software fix to work out how to attack systems that haven’t installed it. The framework, called exploit_agent, analyzes patches, writes attack code and runs it without human intervention, according to The Defiant’s account of Google Threat Intelligence Group’s Nov. 16, 2026, report. Source
Researchers said they found the framework on an exposed server used to prepare attacks, alongside stolen victim data. Automating exploit writing and execution doesn’t establish that every step of an intrusion ran itself. Source
The fix gives attackers somewhere to look
Google says exploit_agent continuously compares web application code before and after patch announcements, looks for potential vulnerabilities in the differences, drafts exploits and automatically runs them. It targets already-disclosed flaws without needing to discover a previously unknown bug. Source
That is the uncomfortable arrangement: the code change that repairs a weakness can also help an attacker locate it. Google argues that comparing versions, release notes and related research may give language models a more manageable task than hunting broadly for unknown vulnerabilities. Source
Google described this as the first time it had identified attackers fully automating the work of finding and exploiting software flaws without a human in the loop. Its earlier reports had covered AI-assisted research into software flaws and agents carrying out parts of live intrusions. Source
One example shows the framework inspecting access controls in Spectra 2.19.25, a WordPress plugin also called Ultimate Addons for Gutenberg. Threadlinqs’ account notes that Google supplied no associated vulnerability identifier and did not say this example had been used in an attack. SourceSource
The speed claim still needs a stopwatch
The concern is a shorter window for defenders to install fixes. How much shorter? The Defiant reports that Google provided no measurement of how long exploit_agent took to turn a patch into an exploit. Source
Google said evidence from victims suggested an interest in French media, entertainment and education websites. Its account does not identify a specific threat group or connect an individual generated exploit to a particular victim. Source
The defensive advice remains decidedly unfuturistic. Mandiant’s guidance, as reported by The Defiant, recommends credentials that expire quickly, limits on account permissions, dividing networks into separate sections and monitoring for suspicious behavior. Source
Comments