Google GTIG: Agents Now Turn Patches Into Exploits Unattended
Google Threat Intelligence Group published its latest AI Threat Tracker on 17 November 2026, written with Mandiant. The headline claim: "In the past quarter, GTIG identified threat actors using LLM frameworks to fully automate vulnerability discovery and exploitation without a human in the loop for the first time."
Three parts of the report matter to people who run AI infrastructure. There is an exploit framework that reads patches. A LiteLLM CVE was used to steal model access. And a North Korea-linked cluster planted backdoors through a coding agent's settings file.
What happened
exploit_agent. GTIG found an exposed staging server hosting a multi-agent framework it calls exploit_agent. It "integrates with multiple frontier and local LLMs" (no model is named) and runs continuous differential analysis of popular web application software. It compares code before and after a patch announcement, works out what the patch fixed, drafts an exploit and runs it automatically. Bundled scripts handled the post-compromise steps: internal reconnaissance, brute-forcing, installing webshells and removing other attackers' webshells. A configuration file carried Russian comments describing LLM API keys "for auto-exploitation (exploit_agent pilot)". The example in the report shows the framework reasoning about an access-control check in a WordPress plugin. GTIG says victim telemetry suggests the operator was interested in French media and entertainment websites and the education sector.
Patch windows are shrinking from both sides. The report also covers two PHP flaws, CVE-2026-109618 and CVE-2026-109619, that were found with LLM-assisted analysis and patched on 21 October. A proof-of-concept video and early patched code were public before the fix shipped, followed by hundreds of proof-of-concept exploits that did not work. Separately, two PaperCut zero-days (CVE-2026-81578 and CVE-2026-82078) were chained in late August.
LiteLLM and LLMJacking. Under privilege escalation GTIG writes that a threat actor exploited "CVE-2026-40217, a sandbox escape in LiteLLM that allows an attacker to elevate limited-access inference keys ("Virtual Keys") into full administrative credentials." It ties this to LLMJacking, meaning stolen access to someone else's model spend. Mandiant says single-day fraudulent charges in some incidents "exceeded $100,000 USD" and that the worst cases reached "$1 million USD before remediation." Stolen access is resold through what GTIG calls LLM Mixers, at $0.10 to $25 per million tokens. More detail, including hardening advice, is in GTIG's 3 November LLMJacking report.
A backdoor in .claude/settings.json. A suspected MIDNIGHT NEPTUNE cluster (GTIG links it to the DPRK) went after cryptocurrency firms in North America and Europe. Its route in was fake recruiter profiles and "ClickFix" lures on fake meeting sites. Once inside, the actor modified .claude/settings.json files in internal code repositories. When a developer started Claude Code, the file automatically triggered a download of a backdoor GTIG calls SOMBERMEME. The actor also created GitHub personal access tokens named after Microsoft Copilot to hide its lateral movement. GTIG does not publish the exact settings content.
The LiteLLM CVE needs a careful read
Google's two reports describe CVE-2026-40217 differently, and neither description matches the advisory exactly.
| Source | What it says CVE-2026-40217 is |
|---|---|
| GitHub advisory GHSA-wxxx-gvqv-xp7p | A sandbox escape in the custom-code guardrail test endpoint, POST /guardrails/test_custom_code. Code runs as root in the default Docker image. "Reaching the endpoint requires a proxy-admin credential" in default configurations. Affects >= 1.81.8, < 1.83.10 |
| GTIG, 17 Nov | A sandbox escape that turns Virtual Keys into administrative credentials |
| GTIG LLMJacking report, 3 Nov | A "user ID rebind vulnerability in the /key/regenerate endpoint" |
The likely explanation is a chain. Two other LiteLLM advisories published on 21 May cover exactly the Virtual-Key-to-admin step: CVE-2026-47101 (an internal_user can create keys with admin-only routes, fixed in 1.83.14) and CVE-2026-47102 (a user can set their own user_role to proxy_admin through /user/update, fixed in 1.83.10). Admin access plus the guardrail sandbox escape gives code execution. Neither advisory mentions /key/regenerate, so I cannot confirm GTIG's description of the endpoint. The fix is the same either way: run LiteLLM 1.83.14 or later. The current PyPI release is 1.110.0.
Zenity reported in June that its honeypots saw 569 requests to the guardrail endpoint between 15 April and 24 May. Each guessed a master key such as sk-1234, which is the value in LiteLLM's own examples.
What changes for engineers
- Your patch-to-exploit window is now machine speed. The framework only needs a public diff. For any internet-facing app, treat "patch released" as "exploit available".
- Your AI gateway is a credential vault. A LiteLLM proxy holds provider keys and issues Virtual Keys. An unpatched gateway turns one leaked low-privilege key into the whole bill.
- Repository config is now a delivery channel.
.claude/settings.jsonis committed and reviewed like code, and it can run commands. An attacker with push access to an internal repo can reach every developer who opens it.
What I ran
I tested what a repository's .claude/settings.json and .mcp.json can run on Claude Code 2.1.327, the current npm release, in a scratch folder with an isolated config directory and a dummy API key. In a claude -p run the repository's SessionStart hook, its apiKeyHelper and its .mcp.json server command all ran. None of them waited for a trust prompt, and they all ran before the dummy key was rejected. --setting-sources user stopped all three. Claude Code's permissions page documents this behaviour; it is not a bug. The full setup, a script that flags these files, and the commands I used are in how to audit a repository's coding agent config for backdoors.
What to do this week
- Upgrade LiteLLM to 1.83.14 or later, rotate the master key off any default value, and block
POST /guardrails/test_custom_codeat your reverse proxy if you do not use it. Keep the admin UI and API off the public internet. - Set spend alerts and per-key budgets on every provider account and gateway key. GTIG puts the worst single-day charges above $100,000, so a daily alert matters more than a monthly one.
- Run the config audit across internal repositories, and require code-owner review for
.claude/,.mcp.json,.cursor/and.vscode/. - Check GitHub tokens with odd names. List fine-grained and classic PATs in your organisation and question any named after Copilot or other tools.
- Shorten patch SLAs for internet-facing apps and subscribe to advisories for every framework and plugin you expose.
The incident is logged in the AI agent security incident log, with the LiteLLM CVEs in its advisory table.
FAQ
What is exploit_agent?
It is GTIG's name for a multi-agent framework it found on an attacker's exposed staging server. It compares software before and after a patch, works out the bug, writes an exploit and runs it without a human. GTIG says this is the first time it has seen discovery to exploitation fully automated.
Which LiteLLM versions are affected by CVE-2026-40217?
The GitHub advisory lists >= 1.81.8, < 1.83.10. Related privilege escalation flaws are fixed in 1.83.10 and 1.83.14, so 1.83.14 or later covers all three.
How did attackers use Claude Code's settings.json?
GTIG says a suspected DPRK cluster modified .claude/settings.json in internal repositories it had already reached, so that starting Claude Code automatically downloaded a backdoor. It did not exploit a Claude Code vulnerability. It used a configuration file that runs commands by design.
Related: LiteLLM 1.107.0 supply chain attack, How to Block Dangerous Commands With Claude Code Hooks and AI vulnerability discovery, offense and defense.
Spot an error or something out of date? Email me and I will fix it and log it on the corrections page.
I'm Ajay Kumar — I build and operate PandaStack, an open-source Firecracker microVM cloud for AI agents, and write about the infrastructure and tooling around it.
Need this kind of infrastructure work? See what I do or email hello@ajayk.sh.
Related
wfx-loot: 4,837 Hijacked GitHub Accounts Pushed a CI Secret Stealer
From 10 November 2026 an attacker used stolen GitHub credentials to probe 61,764 repositories and run a secret stealer in at least 1,806 pipelines. What happened, what I counted on GitHub this morning, and what to check this week.
7 minNov 18, 2026How to Audit a Repo's Coding Agent Config for Backdoors (2026)
Step-by-step: find the hooks, helper commands, env overrides and MCP servers a repository can make Claude Code run, tested on 2.1.327. Includes a jq audit script, a CI gate, safe flags for claude -p, and managed settings that lock it down.
11 minNov 17, 2026GitHub Copilot Can Now Force MCP Servers Through Plugins
On 16 November 2026 GitHub added a managed setting that makes Copilot accept MCP servers, hooks, skills and agents only from plugins, and VS Code set February 2027 to remove its Local agent harness. What changed and what to check.
6 min