All articles
ArticlesNovember 17, 20266 min read

GTIG AI threat tracker: agents turn patches into exploits

Google's GTIG AI threat tracker found exploit_agent, a tool that diffs patches and fires exploits with no human. What it means for your patch window.

GTIG AI threat trackerexploit_agentGoogle Threat IntelligenceAI agentssecuritypatch managementLLMJackingn-day vulnerabilities
GTIG AI threat tracker: agents turn patches into exploits

Google Threat Intelligence Group published its quarterly GTIG AI threat tracker on November 16, 2026, and one finding stands out. For the first time, GTIG says it saw attackers use LLM frameworks to fully automate vulnerability discovery and exploitation, with no human in the loop.

The tool is called exploit_agent. GTIG found it on an exposed staging server. It watches popular web software, compares the code before and after a patch, works out what was fixed, writes an exploit and runs it. Then it moves on to the next one.

I do security at erxes, and I maintain open source code that people self-host. My first thought was that our release notes and public diffs are now an attack feed. That changes how fast patches need to land.

What GTIG found in exploit_agent

According to the full GTIG report on the Google Cloud blog, exploit_agent integrates with multiple frontier and local LLMs. It is set up to do continuous differential analysis of popular web application software: diff the patched and unpatched versions, analyze what changed, draft an exploit, execute it.

A few details from the report:

  • One example in the report shows the framework inspecting an access control check in the Spectra (Ultimate Addons for Gutenberg) WordPress plugin, version 2.19.25.
  • An environment file on the server had Russian comments describing LLM API keys "for auto-exploitation (exploit_agent pilot)".
  • The server also held data exfiltrated from victims, plus scripts for internal recon, brute-forcing, installing webshells and removing other attackers' webshells.
  • Victim telemetry suggested the operator went after French websites in media, entertainment and education.

GTIG's earlier trackers only described partial automation. This quarter it calls the shift a "critical inflection point."

Why n-days are the real target

The part I found most useful is GTIG's explanation of why attackers point these agents at n-days instead of zero-days. A patched bug comes with a small, finite set of material: two code versions, release notes, maybe a researcher write-up or a PoC. That is a perfect job for an LLM. Hunting a zero-day means searching much wider.

The numbers in the report fit that. Overall disclosures and exploits nearly doubled between January and August 2026. Zero-day exploits rose only from a 2025 average of 8 per month to 11 per month in 2026, then spiked to 22 in August.

GTIG points to the PaperCut case as an example. In late August 2026, attackers chained two PaperCut print server flaws (CVE-2026-81578, an auth bypass, and CVE-2026-82078, unsafe dynamic class loading) for unauthenticated remote code execution. Researchers and Huntress produced working exploits with LLMs within a day of disclosure. GreyNoise reported a campaign four days after disclosure where an actor likely used LLMs to build exploits, assemble a target list and hit hundreds of organizations, going from a foothold to domain admin within minutes.

There is a side effect defenders will recognize. After a PHP patch on October 21, 2026, hundreds of claimed PoCs showed up, most of them non-functional and likely LLM-generated. Somebody still had to check each one.

Offensive agent frameworks and LLMJacking

The report also covers the open source offensive frameworks I wrote about in my post on the FBI ARTEX alert. Besides ARTEX, GTIG has seen CAIRN and PentAGI deployed in the wild, and experimentation with CyberStrikeAI and HexStrike AI. CAIRN was used against targets including the US aerospace sector. A customized PentAGI platform was used for recon and scanning against organizations in Asia between late August and early October 2026.

GTIG notes that these tools currently trade stealth for speed. They make a lot of noise and leave code-level traces. GTIG expects that to change as the tools mature.

On the money side, Mandiant responded to several LLMJacking incidents in 2026, where attackers broke into cloud environments to burn the victim's compute and LLM quotas. Single-day fraudulent charges exceeded $100,000, and in the worst cases cumulative abuse reached $1 million before it was stopped. If your API keys for model providers sit in a CI variable or a .env file on a shared box, that line should bother you.

Coding agents as a backdoor path

This is the one I would forward to every engineering team. GTIG describes a suspected DPRK cluster, MIDNIGHT NEPTUNE, that targeted crypto companies in North America and Europe. They used fake recruiter profiles and ClickFix lures for initial access. Then they moved into internal repos and modified .claude/settings.json so that starting Claude Code would download the SOMBERMEME backdoor. They also created GitHub personal access tokens named after Microsoft Copilot to blend in.

Agent config files are code now. Hooks and startup commands in those files run on developer machines with developer credentials. Most teams don't review them the way they review a package.json script, and they should. I covered a related angle in securing AI agents with Uber's framework.

What I'm changing on my side

Here is my short list after reading the report:

  1. Treat security patches as a race. For internet-facing apps, plugins and self-hosted tools, the clock starts when the fix is public, including preview channels and public commits. Have a path to deploy in hours.
  2. Watch your public diffs. If you ship open source, a security fix in a public commit is readable by an agent before your users upgrade. Coordinate disclosure, batch the release and push the advisory at the same time as the fix.
  3. Review agent config like CI config. Add .claude/, .cursor/, MCP configs and similar files to CODEOWNERS and require review on changes to hooks or startup commands.
  4. Put hard limits on LLM spend. Set spend caps and alerts with each model provider, scope keys per service, and rotate anything that ever touched a shared machine.
  5. Expect noise. These frameworks are loud today. Rate limits, WAF rules on admin routes and alerting on bursts of failed auth still catch a lot.

FAQ

What is exploit_agent?

It is the name GTIG gave a multi-agent framework found on an exposed attacker staging server. It works with multiple LLMs, diffs patched and unpatched versions of popular web apps, and automatically drafts and runs exploits.

Is this the first fully autonomous AI exploitation GTIG has seen?

Yes, according to the report. GTIG says that in the past quarter it identified threat actors using LLM frameworks to fully automate vulnerability discovery and exploitation without a human in the loop for the first time. Earlier trackers described partial automation.

What is LLMJacking?

Stealing LLM credentials or cloud resources from a victim and using or reselling that access. Mandiant saw single-day fraudulent charges above $100,000 in some 2026 incidents.

How fast should I patch now?

GTIG's data and the PaperCut case point to working exploits within a day of disclosure and mass exploitation within days. For exposed systems I would plan for hours.

My bet for the next tracker: the noise goes down. GTIG's own timeline puts fully agentic attacks with no human in the loop within the next one to six months, and quieter tooling is the obvious next step.


Sources:

Cover image: "Evolution of threat actor use of LLMs" timeline from Google Threat Intelligence Group.

Related articles