Connect with us

NEWS

Google Finds an AI Agent That Feeds on Public Patches

Google found exploit_agent, an AI that turns public patches into attacks because a fix is a finite map for models.

Published

on

Google Threat Intelligence Group said Nov. 16 it found an AI framework that reads public software patches, writes exploits and fires them with no person in the loop. GTIG named the kit exploit_agent after finding it on an exposed staging server, wired to frontier models and to copies run on local machines.

The same box held data stolen from victims, plus scripts for password guessing, webshells and wiping rival webshells. A published fix is now a finite map. That is the part that changed.

An Exposed Server Held the Patch-Reading Agent

Aurora Blum, Kelli Vanderlee, Jess Xia and David Mainor wrote the quarterly AI Threat Tracker with Mandiant. In the past quarter, they said, GTIG identified threat actors using LLM frameworks to fully automate vulnerability discovery and exploitation without a human in the loop for the first time.

The earlier tracker, published in September, had described partly automated scanning and said GTIG had not seen that full loop used against real targets. exploit_agent is built to keep comparing popular web apps before and after each patch note, then draft an exploit and run it.

One figure in the report shows the agent inspecting an access-control check in version 2.19.25 of Spectra, the Ultimate Addons for Gutenberg WordPress plugin. GTIG did not say that plugin was popped in the wild. Victim telemetry on the server pointed at French media, entertainment and education sites.

A configuration file carried Russian comments. GTIG’s translation labels the API keys “for auto-exploitation (exploit_agent pilot)” and tells the operator, “Enter the keys yourself, right here, not in the chat.” GTIG named no group and no state. The language is a clue, not a charge sheet.

Google’s public note on the tracker put a harder line on the calendar. It said the era of autonomous attacks has arrived, and that defending against machine-speed threats at human speed is not feasible.

Why Patched Bugs Beat Unknown Ones for AI

Unknown bugs still matter. GTIG’s own counts show they are no longer the efficient meal. Overall disclosures and exploits nearly doubled between January and August 2026, while zero-day exploits only moved from a 2025 average of 8 a month to 11 a month in 2026, then jumped to 22 in August.

A zero-day hunt means casting a wide net. A patched n-day comes with the old tree, the new tree, the release notes, the write-ups and, often, a proof of concept. That is a closed set. Models are good at closed sets.

Threat actors may find it more efficient to use LLMs to rapidly weaponize patched n-day vulnerabilities, as they can focus research on a finite set of source materials: comparing source code versions (including from pre-release or preview channels), and examining release notes, security research, and POC code related to specific flaws.

Google Threat Intelligence Group, AI Threat Tracker, Nov. 16, 2026

The junk is part of the same loop. GTIG said the lower bar has produced a wave of LLM-written exploits that are claimed to work and do not, which dumps extra checking on defenders. Speed and sludge arrived together.

Four Hours From an Empty Workspace to a Real Breach

PaperCut NG and MF print servers showed how short the gap can get once a fix is public. Huntress saw exploitation in two customer environments, including activity on Aug. 26 that lasted under two minutes, and reproduced a pre-authentication remote code execution chain against a stock PaperCut NG 25.0.11.75758 server.

PaperCut published an urgent advisory on Aug. 27 and emergency patches on Aug. 28 for CVE-2026-81578, an access-control flaw, and CVE-2026-82078, unsafe class loading. Chained, they give an unauthenticated attacker code execution on the application server. GTIG said researchers and Huntress used large language models to produce working exploits within a day of the disclosure.

THE PAPERCUT CLOCK IN LATE AUGUST

  1. Aug. 26, 2026: Huntress records exploitation in a customer environment lasting under two minutes.
  2. Aug. 27, 2026: PaperCut warns that NG and MF are under active attack and treats all versions as in scope.
  3. Aug. 28, 2026: Emergency patches ship; GTIG says working AI-assisted exploits exist within a day of disclosure.
  4. Aug. 31, 2026: GreyNoise times an actor from an empty workspace to a live shell in under four hours, then 11 organizations in 26 seconds.

GreyNoise’s PaperCut mass exploitation timeline puts that campaign four days after the advisory. The actor used OpenAI’s Codex harness with a DeepSeek model. GreyNoise counted 440 compromised instances at 395 organizations in 48 countries, with 204 of those instances in education and domain-admin access in 12 cases. First domain admin came two hours after the first real shell. In one U.S. high school the path from access to domain admin took seven minutes.

The agents still needed an open server. Huntress said 47 percent of the about 2,500 PaperCut installs it tracks were on v23 or older, a branch with no patch at the time, and urged operators to pull the application server off the public internet even if they could not upgrade that day. A weekly change window is now longer than the clock GreyNoise published.

After PHP Shipped Fixes, GitHub Filled With Dead Exploits

PHP, the language behind roughly seven of ten websites, closed CVE-2026-109618 and CVE-2026-109619 on Oct. 21, 2026. GTIG said a proof-of-concept video and early patched code were public before those official builds, which let researchers and attackers compare versions. Hundreds of claimed proofs of concept followed. Many did not work. GTIG said they were likely LLM-generated.

WHAT THE PHP FIX HANDED ATTACKERS

  • The map: Orca Security said the repair changed 300 files and added 127 regression tests written to trigger the bugs.
  • The builds: Fixes landed in PHP 8.2.35, 8.3.36, 8.4.27 and 8.5.12; Orca verified the repair stops the attack it built.
  • The hours: Roi Nisimi’s team built our own exploit with AI from the published fix, then ran commands on a standard PHP server and on WordPress within hours.
  • The sludge: GTIG said the flood of broken proofs created extra verification work for defenders even as a few real tools circulated.

Orca also found that a stranger’s GitHub tool, run unchanged, took over an unrepaired test server. Guardrails on big hosted models blocked a full chain until the team switched agents. Local models with fewer brakes are already in that market. The useful warning is the setup, not a smarter chatbot: anyone can diff the patched tree against the old one and leave a model to grind through version quirks.

Live Intrusions Already Run on Open Agent Frameworks

exploit_agent is one kit on one exposed host. GTIG and Mandiant also watched open pentest frameworks used in live jobs by financially motivated crews and by suspected espionage actors with a PRC nexus, including signs that a PRC-nexus group was building its own automated attack framework.

OPEN AGENT KITS GTIG TIED TO REAL JOBS

Framework How it is built What GTIG observed
CAIRN OODA loop with specialist agents on a shared blackboard Intrusion activity against targets including U.S. aerospace
PentAGI Open autonomous kit, customized as an orchestration platform Recon and vulnerability scanning against Asian organizations from late August to early October
CyberStrikeAI Go backend tied to 100-plus security tools and YAML recipes Experimentation, not a named sector campaign in this tracker
HexStrike AI MCP harness wrapping about 150 tools and more than 12 sub-agents Experimentation alongside CyberStrikeAI
ARTEX Widely discussed open pentest stack Google Security Operations shipped behavioral and volumetric rules to spot it

GTIG wrote that it is highly likely these frameworks will keep showing up in campaigns worldwide. They still hit familiar holes, stolen passwords and unpatched boxes, only faster. Mandiant published a hardening guide aimed at that speed, and Google Security Operations added ARTEX detections. A louder kit is easier to write rules for than a quiet one.

What LLMJacking Costs When a Key Leaks

The same tracker follows a second business: stealing cloud and model keys, then burning the victim’s quota or reselling access through underground “LLM Mixers.” Mandiant spent 2026 on several of those jobs, a pattern GTIG calls LLMJacking.

BILLS AND BACKDOORS IN THE SAME QUARTER

  • One-day burn: Fraudulent charges topped $100,000 USD in a single day on hijacked model access.
  • Before cutoff: In the worst cases Mandiant saw, cumulative abuse reached $1 million USD before access was cut.
  • LiteLLM: One path used CVE-2026-40217, a sandbox escape that turns limited inference keys into admin credentials.
  • Claude Code: A suspected North Korean group edited.claude/settings.json inside companies so each launch pulled the SOMBERMEME backdoor.

Google said it disabled accounts and assets tied to Gemini misuse and that DeepMind used the cases to strengthen classifiers. The stolen-compute racket and the patch-reading agent share a habit. They live on credentials and diffs that organizations already publish or leave in logs, not on magic zero-days.

Forensic Noise Is the Window Defenders Still Have

GTIG’s own limit is easy to miss under the first-time headline. These kits buy speed by spraying traffic, leaning on known tool marks and leaving code-level traces. That is a detection gift, and GTIG does not pretend it lasts.

They currently trade stealth for velocity. Unrestricted autonomous execution generates forensic traffic, relies on recognizable identifiers, and leaves behind distinct code-level traces.

Google Threat Intelligence Group, AI Threat Tracker, Nov. 16, 2026

As the models improve, GTIG expects less human steering and quieter tradecraft. Until then the fight is still the unpatched CMS, the print server on the public internet, the LLM gateway on port 4000 with a default key. exploit_agent was found because someone left the staging server open. The next copy may not.

Harry is the editor and lead writer of STUDIO ONE NETWORKS, an independent title he owns and runs himself. Ten years in journalism, reporting first and editing later, taught him that entertainment and business are one beat seen from two sides: a box office figure is a company number, a streaming deal is a contract, a casting rumour is not a story until someone puts their name to it. He works from the record, whether that is a distributor's statement, a licensing agreement, an interview transcript or a set of published ratings, and checks every number against it before publication. The same rule holds for the rest of the site, which covers news, technology, science, sports, lifestyle, travel, auto and gaming for an audience spread across the world. When he gets something wrong, the article is corrected and the change is noted and dated, under a corrections policy anyone can read. Reader mail is answered by him at support@studioonenetworks.com.

Click to comment

Leave a Reply

Your email address will not be published. Required fields are marked *

Trending