Google says attackers built an AI framework to turn patches into exploits without human involvement
Google’s Threat Intelligence Group said Monday that attackers built an AI framework meant to turn software patches into working exploits automatically. Google called this the first time it has seen LLM systems used to fully automate vulnerability discovery and exploitation “without a human in the loop.” Researchers found the tooling on an exposed staging server next to stolen victim data and post-compromise scripts; the report does not establish that the framework itself caused those compromises.
In its configuration the framework is named exploit_agent. It wires together multiple frontier and local large language models to compare code before and after patches, flag possible weaknesses, draft exploits, and run them without human intervention. Google named no specific flaw that it exploited, no victim count, and no measured patch-to-exploit time. Merritt Baer, a former deputy CISO at AWS and an advisor to Upwind and G2i, said attackers had not invented a new class of bug but had automated reading the patch, seeing what it fixed, and writing the exploit—work that might take a skilled person days.
The operator’s config file includes Russian-language comments calling exploit_agent a pilot and telling the user to paste LLM API keys into the file. Google does not attribute the framework to a named group or country. Victim telemetry on the server suggested interest in French media, entertainment, and education sites. Google published a code excerpt in which the framework inspects access controls in version 2.19.25 of the Spectra WordPress plugin, without tying that look to a successful exploit. Earlier Google reporting covered partially automated scanning and live use of other AI frameworks, including CAIRN and PentAGI; what sets exploit_agent apart is its design around patches. Baer’s defensive point was ordinary hygiene: know what is on the internet, patch faster than last year, and watch for exploitation, not only for vulnerabilities.
Continue in the app — vote & join in ➔