Independent · Source-cited · No sponsored rankingsSources linked in every article
AIToolRanked
AI News · 5 min read

Google GTIG: Claude Code settings backdoor and LLMJacking

On November 16, 2026, GTIG reported a Claude Code settings backdoor and LLMJacking bills past $100,000 a day. Google described no Claude Code flaw.

RA
· Founder, AIToolRanked
A code editor with colored bars and one red line, a dotted line to a download icon, a cluster of agent nodes, and a warning shield
On this page

What GTIG reported on November 16

On November 16, 2026, Aurora Blum, Kelli Vanderlee, Jess Xia, and David Mainor published the GTIG AI Threat Tracker ↗. CyberScoop ↗ and BleepingComputer ↗ covered it the same day. The picture above is ours. We did not reproduce these intrusions. The report publishes no SOMBERMEME hashes, and this page adds none.

Google does not describe a vulnerability in Claude Code itself. The attackers already had write access to the repositories.

In its section on Democratic People's Republic of Korea actors, GTIG describes a suspected MIDNIGHT NEPTUNE cluster. BleepingComputer ↗ names the activity cluster UNC6836, with suspected links to that name. The targets were cryptocurrency organizations in North America and Europe. For initial access, GTIG says the actor used fake recruiter profiles whose images carried SynthID watermarks, and sent people to fake meeting platforms with "ClickFix" lures that asked them to run "diagnostic fix" scripts. After the compromise, the actor modified .claude/settings.json in internal code repositories so that starting Claude Code downloaded the SOMBERMEME backdoor, and created GitHub personal access tokens named after Microsoft Copilot.

The pipeline, the patch window, and the bills

GTIG says threat actors used large language model frameworks "to fully automate vulnerability discovery and exploitation without a human in the loop for the first time." The case is an exposed staging server running a multi-agent framework GTIG calls exploit_agent, wired to multiple frontier and local models. It compares popular web application code before and after a patch announcement, picks out likely flaws, drafts exploits, and executes them with no person in the loop. The server also held scripts for reconnaissance, brute-forcing, installing webshells, and removing other attackers' webshells. The targets on that server were French websites in media, entertainment, and education. Russian-language comments sat in the configuration. GTIG did not name a group or a state. CyberScoop ↗ reports the same limit.

GTIG's count of exploited zero-days is a 2025 average of 8 a month, 11 a month in 2026, then 22 in August 2026. Overall disclosures and exploits nearly doubled between January and August 2026.

After the PHP patches on October 21, 2026, for CVE-2026-109618 and CVE-2026-109619, hundreds of claimed proof-of-concept exploits appeared. GTIG says they did not work, and describes them as likely generated with large language models.

The report's summary names the frameworks ARTEX, CAIRN, CyberStrikeAI, and PentAGI. The body also describes HexStrike AI, which BleepingComputer ↗ lists as well. GTIG says CAIRN was used against targets including the aerospace sector in the United States. CyberScoop and BleepingComputer describe that activity as a U.S. aerospace organization.

BleepingComputer reports the researchers' estimate in two horizons: unsophisticated attacks needing "little to no human interaction" within three months, and "sophisticated, fully autonomous" agentic attacks likely within six months.

On LLMJacking, Mandiant's 2026 incident-response cases include single-day fraudulent charges above $100,000. The most severe cumulative abuse reached $1 million before remediation. Initial access often came from credentials taken out of public GitHub repositories and from session hijacking. One actor exploited CVE-2026-40217, which GTIG describes as a LiteLLM sandbox escape that elevates limited "Virtual Keys" into full administrative credentials. A botnet generated millions of LLM API calls in about 24 hours. GTIG says "LLM Mixers" resell stolen access, with advertised prices from $0.10 to $25 per million tokens.

What people are saying

  • "I'd put .claude/ under CODEOWNERS and review it like the build script" — token budget ↗ (@token_budget)

  • "No bug in the model or in Claude Code was needed." — NULLchimp ↗ (@nullchimp)

  • "open .claude/settings.json in any repo you didn't write and read the hooks before you start it there." — Noor Anand ↗ (@noorodivergent)

  • "Please, there is no way that's not been going on for months." — DQ52, a reader comment on the BleepingComputer article ↗

Our take

If you run Claude Code, Codex, or Cursor, the agent config is a code-execution surface. Anthropic's hooks documentation ↗ says .claude/settings.json is project-scoped and can be committed to the repo. Hooks are shell commands and HTTP endpoints that run on their own, including at SessionStart. Treat .claude/, hooks, and MCP configs like a CI workflow: CODEOWNERS, a reviewed diff, and branch protection.

Open /hooks before you start Claude Code in a repository you did not write. The menu is read-only and labels each hook's source: user, project, local, or a plugin. In an interactive session, Claude Code holds project hooks until you accept the workspace trust dialog. A claude -p or SDK session treats the folder as trusted, so a committed hook runs without that dialog.

A switch in your own user settings is not enough. Anthropic documents "disableAllHooks": true as the way to turn hooks off. A project file that sets "disableAllHooks": false overrides a true in user settings. For one run, whatever the project file says, pass --settings '{"disableAllHooks": true}'. Hooks set by an administrator in managed policy are a separate tier, and a user or project flag does not turn those off. The same month, the MadPot note covers ARTEX agents and stolen API keys, including on LiteLLM.

Teams that hold LLM API keys should expect a stolen key to show up on the bill the same day. Set billing alerts and quotas. Keep keys out of repositories. Apply the vendor fix for CVE-2026-40217 on any LiteLLM host you run, then rotate the keys. For internet-facing web apps, turn on automatic updates. The non-functional PHP proofs are noise. The working pipeline is the reason to patch.

Google did not describe a vulnerability in Claude Code itself, or in the model. NULLchimp's sentence matches the report: write access was enough. The earlier settings.local.json leak was secrets in an approval file. The Cursor incident was a vendor key. California's lawsuit is a pleading about research sandboxes, not this intrusion. Compare the product in the Claude Code review. Review tooling is the AI code review tools guide.

More from the news desk · About Rai Ansar