Google finds AI agent that turns security patches into exploits

Google's threat researchers say an attacker's framework found flaws and exploited them with no human involved, a first in their tracking.

Illustration of a robotic arm comparing two glowing sheets of code on a dark desk, with a padlock icon splitting open between them and red warning lights on server racks in the background. Image: AgenticFeed illustration

Google Threat Intelligence Group said on Monday that it has seen attackers use an AI framework to find software vulnerabilities and exploit them with no human in the loop, the first time it has observed a fully automated pipeline of that kind.

The finding appears in GTIG's quarterly AI Threat Tracker, published with Mandiant under the title "Agentic Threat Activity Crosses a Significant Threshold." In its previous edition two months ago, GTIG said it had not observed fully autonomous attack pipelines deployed against targets in the wild, SiliconANGLE reported.

How exploit_agent works

GTIG said it found the framework, which it calls exploit_agent, on an exposed staging server. The system plugs into multiple frontier and locally hosted large language models and runs continuous differential analysis of popular web application software. In practice, it compares code before and after a patch is announced, works out which flaw the patch fixed, drafts an exploit and then runs it automatically.

The same server held data stolen from victims along with scripts for post-compromise work. According to GTIG, those scripts covered internal reconnaissance, brute-forcing and installing webshells, plus removing webshells left by other attackers. Victim telemetry suggested the operator focused on French websites in media and entertainment as well as education.

GTIG said the tooling used Russian-language comments. One line in a configuration file, translated by GTIG, tells the user to enter the LLM API keys "right here, not in the chat." The Register described the operator as suspected to be Russian-speaking.

GTIG's report shows an example of the framework inspecting access-control code in version 2.19.25 of Spectra, a WordPress plugin also known as Ultimate Addons for Gutenberg.

Why patched bugs are the target

GTIG argues that attackers may find it more efficient to point LLMs at n-day vulnerabilities, flaws that already have a patch, than to hunt for zero-days. A patch gives a model a finite set of material to study, including code diffs and release notes. The result, the report says, is a shrinking window between when a fix ships and when organizations face working attacks against it.

The report's own numbers point the same way. GTIG said overall vulnerability disclosures and exploits nearly doubled between January and August 2026, while zero-day exploits rose only from a 2025 average of 8 per month to 11 per month in 2026, before jumping to 22 in August.

It also cites outside cases. In late August, attackers exploited two chained PaperCut flaws, and GTIG said Huntress and other researchers used LLMs to build working exploits within a day of disclosure. Greynoise reported a campaign four days after disclosure in which an attacker likely used LLMs to build exploits and hit hundreds of organizations, moving from a foothold to domain administrator within minutes.

According to GTIG, cheap exploit generation also creates work for defenders. After PHP patches on Oct. 21, the report said, hundreds of claimed exploits appeared that did not work, likely produced with LLM help, and security teams had to verify or debunk each one.

Open-source agent frameworks in live intrusions

Beyond exploit_agent, GTIG said multiple threat actors are now using open-source agentic penetration-testing tools during real intrusions. It named ARTEX, CAIRN, CyberStrikeAI and PentAGI, and described HexStrike AI as an offensive harness built on the Model Context Protocol that connects LLM clients to more than 150 security tools. GTIG said it saw CAIRN used against targets including the US aerospace sector, and tracked a customized PentAGI platform scanning organizations in Asia between late August and early October.

GTIG noted that these frameworks currently trade stealth for speed, leaving forensic traffic and code-level traces that raise the odds of detection. The group said it expects that to change as the tools and models improve, with less human involvement and quieter operations.

Google used the report to promote its own defenses, including Google AI Threat Defense and its CodeMender patching agent, and said Google Security Operations has deployed detection rules for ARTEX.

Source: Google Threat Intelligence Group, “GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold”

Additional sources: Google, “The era of autonomous attacks has arrived”; The Register, “Google: AI agent turned a security patch into a working exploit with no human help”; SiliconANGLE, “Google says an AI framework turned security patches into exploits with no human in the loop”

Get the AgenticFeed newsletter

A free, regular roundup of the agent stories that matter. No spam, and you can unsubscribe at any time.

Want to reach readers following AI agents? Advertise with AgenticFeed

Next event · 19 Nov, 2026

MCP Dev Summit Nairobi

The Linux Foundation’s Nov. 19 and 20 summit brings Model Context Protocol developers to Kenya, with keynotes from OpenAI, GitHub and the Gates Foundation.

The AgenticFeed newsletter

A free, regular roundup of the agent stories that matter. No spam.