Google Finds an Attacker's Staging Server Running a Patch-to-Exploit Pipeline With No Human in the Loop

Google Threat Intelligence Group's November 16, 2026 AI Threat Tracker describes exploit_agent, a multi-agent framework that diffs patched web software, drafts exploits and fires them automatically. The report also tracks open-source offensive agent frameworks in live intrusions, LLMJacking bills that reached $1 million, and a backdoor planted through Claude Code settings files.

A row of unattended mechanical arms in a dark carmine and slate-green workshop studies a repaired iron hinge under a lamp and forges a crowbar further down the line beside empty stools.
A row of unattended mechanical arms in a dark carmine and slate-green workshop studies a repaired iron hinge under a lamp and forges a crowbar further down the line beside empty stools.

Attackers have started letting software agents turn security patches into working exploits without a person signing off on each step. That is the headline finding of the latest AI Threat Tracker from Google Threat Intelligence Group (GTIG) and Mandiant, published on November 16, 2026. GTIG says that in the past quarter it saw threat actors use LLM frameworks “to fully automate vulnerability discovery and exploitation without a human in the loop for the first time.”

The evidence came from an exposed staging server. On it, GTIG found a multi-agent framework the operator called exploit_agent, wired to several frontier and local language models. According to the report, the framework runs continuous differential analysis of popular web application software: it compares code before and after a patch is announced, works out what flaw the patch fixed, drafts an exploit, and executes it automatically.

What sat on the server

The server held more than the framework. GTIG lists data taken from victims and scripts for post-compromise work, including internal reconnaissance, brute-forcing, installing webshells and removing webshells left by other attackers. One figure in the report shows the framework inspecting access-control code in version 2.19.25 of the Spectra (Ultimate Addons for Gutenberg) WordPress plugin.

A configuration file carried Russian-language comments, which GTIG translates as “LLM API keys for auto-exploitation (exploit_agent pilot)” and an instruction to enter the keys directly into the file rather than into the chat. Victim telemetry suggested interest in French websites in media and entertainment and in the education sector. The report does not name the operator, the victims or specific CVEs tied to the framework.

Off-the-shelf offensive agents in real intrusions

Beyond the custom pipeline, GTIG says several actors are now running open-source agentic penetration-testing frameworks during live intrusions. Alongside ARTEX, the report names four:

  • CAIRN, which uses an observe-orient-decide-act loop with agents sharing a common “blackboard.” GTIG saw it used against targets including the U.S. aerospace sector.
  • CyberStrikeAI, a Go-based orchestration server that connects commercial and open-weight models to more than 100 security tools and YAML attack recipes.
  • HexStrike AI, an MCP-integrated harness exposing more than 150 offensive tools through more than 12 specialised sub-agents.
  • PentAGI, a customised version of which one actor used between late August and early October 2026 to scan a wide range of organisations in Asia.

GTIG adds that a PRC-nexus group appeared to be building its own automated attack framework. Its assessment is that these tools currently “trade stealth for velocity”: they generate noisy traffic and leave code-level traces that help defenders, but they still reach their objectives.

A shrinking patch window

The report places exploit_agent inside a broader trend. GTIG counts disclosures and exploits nearly doubling between January and August 2026. Zero-day exploitation rose more slowly, from a 2025 average of 8 per month to 11 per month in 2026, before jumping to 22 in August.

Two cases illustrate the speed. For the PaperCut zero-days disclosed in late August, the report notes that researchers generated working exploits with LLMs within a day, and that a widespread campaign followed four days after disclosure. Around the PHP flaws patched on October 21, 2026, early patched code appeared before the official fix, and hundreds of non-functional, likely LLM-generated proof-of-concept exploits followed, each one needing to be checked or debunked.

Stolen keys, resold models

A large part of the report covers LLMJacking, where attackers use stolen cloud credentials to run up model and GPU usage. Mandiant says single-day fraudulent charges in some 2026 incidents exceeded $100,000, and in the worst cases cumulative abuse reached $1 million before it was stopped. In one intrusion, an attacker exploited CVE-2026-40217, a LiteLLM sandbox escape that turns limited “Virtual Keys” into full admin credentials. Others used a “burn and rotate” approach, creating and deleting short-lived service accounts to get around rate limits.

That access is resold through what GTIG calls LLM Mixers, storefronts that route buyers through pools of stolen credentials. Advertised prices ranged from $0.10 to $25 per million tokens.

Coding-agent settings as a backdoor

The espionage section describes a suspected North Korean cluster, MIDNIGHT NEPTUNE, targeting cryptocurrency firms in North America and Europe. After gaining access through fake recruiter profiles and “ClickFix” lures, the actor edited .claude/settings.json files in internal repositories so that starting Claude Code would download a backdoor GTIG tracks as SOMBERMEME. The actor also named GitHub personal access tokens after Microsoft Copilot to blend in.

For defenders, Google says Security Operations now ships behavioural and volumetric rule packs for ARTEX, and Mandiant has published a hardening guide aimed at the speed of LLM-enabled tooling. The company is also pitching Google AI Threat Defense, which combines Gemini and other models with Wiz prioritisation, CodeMender patching and Mandiant intelligence.

Topics
  • #Google Threat Intelligence
  • #Mandiant
  • #Offensive AI
  • #LLMJacking
  • #Vulnerability Exploitation
Raj M

Author

Raj M

Contributor

AI Systems Architect is a seasoned technology leader with over 15 years of experience in the IT industry working with Fortune 500 companies. With a solid foundation in multi-agent systems, open-source LLM infrastructure, and enterprise deployment, he excels at building scalable production-grade AI platforms.