Thursday 19 November 2026
Security

Google details AI framework that turns patches into exploits

An exposed staging server ran exploit_agent, which compared patched and unpatched code and ran attacks unattended. Victim data pointed at French media, entertainment and education sites.

Googleplex headquarters building in Mountain View, California
Google's headquarters, the Googleplex, in Mountain View, California. Photograph: Asoundd/Wikimedia Commons (CC BY-SA 4.0)

Google Threat Intelligence Group said in its quarterly AI threat tracker Monday that an AI framework called exploit_agent, linked to a likely Russian-speaking operator, produced working exploits from security patches without human involvement. Google described the activity as the first fully autonomous vulnerability discovery and exploitation it has observed.

The exposed staging server hosted exploit_agent, which works with several frontier and local LLMs, The Register reported. The tool continuously compares popular web application code before and after fixes, identifies probable flaws, drafts exploits, and executes them automatically. The same server contained stolen victim data and several post-compromise scripts for internal reconnaissance, brute-forcing, installing web shells, and removing web shells left by other intruders.

Victim data on the server indicates the operator focused on French websites in media, entertainment, and education, according to The Register. A configuration file comment written in Russian tells the operator to put API keys "right here, not in the chat," SiliconANGLE noted.

SiliconANGLE said that two months earlier, the tracker had not observed fully autonomous attack pipelines used against live targets. Google titled the report "Agentic Threat Activity Crosses a Significant Threshold." Because patched flaws offer limited source material, GTIG believes attackers may get more value from pointing LLMs at n-day flaws than from hunting zero-days.

The report gives an example: two PaperCut print management server flaws, CVE-2026-81578 and CVE-2026-82078, were chained and used as zero-days in late August. Security researchers and Huntress Labs used LLMs to create working exploits within a day of the flaws being disclosed. A GreyNoise campaign four days later likely used LLMs to build exploits and hit hundreds of organizations, and in one case a foothold reached domain administrator within minutes.

The report also documents use and experimentation with open source agentic frameworks, including ARTEX, CAIRN, CyberStrikeAI, HexStrike AI, and PentAGI. CAIRN appeared in an intrusion against the U.S. aerospace sector, and a customized PentAGI platform scanned organizations in Asia between late August and early October. GTIG said it saw financially motivated actors and suspected PRC-nexus espionage actors experimenting with these tools. Google researchers said it is "highly likely" that threat actors will keep using these frameworks in global campaigns.

The findings follow an OpenAI claim reported by Five Nines earlier this month: an automated pipeline built around an unreleased cyber model produced working local privilege escalation exploits for 948 Linux kernel flaws in one October run.

For defenders, Google Security Operations has put in place behavioral and volumetric rulepacks that detect ARTEX, and Mandiant released a hardening guide for LLM-enabled offensive tooling. The report notes that unrestricted autonomous execution leaves forensic traffic, relies on recognizable identifiers, and produces distinct code-level traces that defenders can use to spot these tools despite their speed.

Sources: theregister, SiliconANGLE, Google Cloud, Five Nines

Noor Haddad

Noor Haddad covers attacks, breach disclosures and exploited vulnerabilities for Five Nines. Noor writes from CISA and vendor advisories, SEC disclosures and published reporting.

Corrections. Report an error to corrections@thefivenines.com. We list substantive corrections on the corrections page.