Google Says Attackers Rewrote Claude Code Settings Files to Pull a Backdoor, and an AI Agent Was Turning Patches Into Exploits. Check Your .claude/settings.json Today

If your developers use Claude Code, open every .claude/settings.json in your repositories today and read it. Look for a "hooks" block, and for any command in it that downloads or runs something: curl, wget, Invoke-WebRequest, a bash -c or powershell one-liner, node -e, python -c. Then ask who added it. Run git log --format='%h %an %ae %ad %s' -- .claude/settings.json in each repo and check that every change came from someone you know, through a reviewed pull request. Do the same for .claude/settings.local.json, .mcp.json and .vscode/tasks.json, which can also run commands when a project is opened.
The reason: Google Threat Intelligence Group (GTIG) published research on November 16, 2026 saying a suspected North Korean cluster it calls MIDNIGHT NEPTUNE broke into cryptocurrency companies in North America and Europe, then edited Claude Code configuration files inside the victims' own code repositories. In GTIG's words, the actor was "modifying .claude/settings.json configuration files so that when users started up Claude Code, it would automatically trigger the download of the SOMBERMEME backdoor."
That is the same trick a Shai-Hulud worm variant used in October: it committed a planted .claude/settings.json into every repository a stolen GitHub token could reach, so the payload ran again whenever a developer opened the project in Claude Code. Two unrelated campaigns have now used the file. Treat it like a CI workflow file: it runs code, so it needs an owner and a review.
What to do today
Claude Code and AI coding agent users:
Read every .claude/settings.json and .claude/settings.local.json you have, in each repo and in your home folder (~/.claude/settings.json). Any hook you did not write is a finding. A hook that fetches something from the internet when a session starts is the shape GTIG describes.
Check the history of those files with git log, as above. A change that arrived in a commit nobody reviewed, or from an account that does not normally touch tooling, is the signal. The Shai-Hulud variant committed under a name made to look like Claude ([email protected], message "chore: update dependencies"); GTIG does not say what MIDNIGHT NEPTUNE's commits looked like.
Protect the paths. Add .claude/, .mcp.json, .vscode/ and .github/workflows/ to CODEOWNERS and require review on the default branch, so a stolen developer token cannot slip a hook in unreviewed.
If you find a planted hook, take it out before you revoke anything. The October Shai-Hulud variant (tensorlake 0.5.144 on npm) installed a watcher that wiped the home directory once its stolen GitHub token was revoked, so its cleanup order was: remove the watcher, then revoke. GTIG does not say SOMBERMEME does this, but clean first, then rotate.
A free check that does step 1 for you. Our check-repo tool, part of Membrane, our pre-install checks for agents and packages, reads a repository's .claude/settings.json, .mcp.json, .cursor and .vscode/tasks.json, plus package install hooks and dev container commands, and tells you what will run when an agent or editor opens the folder. We tested it today against a SessionStart hook that downloads a script and pipes it to a shell, the shape GTIG describes: it returned "block" (rule RLS-003, critical). A harmless hook that runs a local script returns "review", because any hook you did not write deserves a look. It is free and needs no account: call check-repo from the Jeevesus MCP server (docs.dugganusa.com/agents/jeevesus/), or run the DugganUSA Checks GitHub Action on every pull request so a planted hook gets flagged before anyone merges it. It is static: nothing in your repo is fetched, run or stored. "No findings" means none of its patterns matched, not that the repo is safe.
GitHub organization owners:
Review the personal access tokens that can reach your organization, and ask each developer to review their own token list. GTIG says the actor "generat[ed] GitHub Personal Access Tokens (PATs) named after Microsoft Copilot" to hide lateral movement. A token named after Copilot, or any AI tool, that nobody remembers creating is worth a phone call.
Everyone:
Tell staff that a recruiter or meeting page asking them to run a "diagnostic fix" script is the attack. GTIG says MIDNIGHT NEPTUNE's fake recruiter profiles led targets to fake meeting sites with ClickFix lures, and the scripts they ran started the intrusion.
Shrink your patch window for internet-facing web software. GTIG found an attacker framework that diffs patches and builds exploits automatically (below). If an AI agent is reading the patch the day it ships, the gap between "patch released" and "exploit in the wild" is the time it takes you to install it.
If you run an AI gateway, rotate the provider keys it holds and check provider billing. GTIG says stolen model credentials are being resold through underground proxy services, and it saw an attacker exploit CVE-2026-40217 in LiteLLM to turn a limited "virtual key" into full admin credentials. We covered the gateway-key theft side yesterday in "AI Agents Stole AI Keys From Exposed LiteLLM, Dify and One API Servers."
What GTIG says is confirmed
The report is "Agentic threat activity crosses a significant threshold," by Aurora Blum, Kelli Vanderlee, Jess Xia and David Mainor of GTIG, dated November 16, 2026. Our quotes come from reading the page itself; The Hacker News article was only how we found it.
An agent that turns patches into exploits. GTIG "identified an exposed staging server hosting a multi-agent vulnerability research and weaponization framework, dubbed exploit_agent, that integrates with multiple frontier and local LLMs." GTIG says the framework "compares code before and after patch announcements, analyzes findings to identify potential vulnerabilities, drafts exploits, and automatically executes them without human intervention." The server also held data taken from victims and scripts for internal reconnaissance, brute-forcing, installing webshells "and removing other attackers' webshells." GTIG says victim telemetry on the server pointed to French websites in media and entertainment and in education. Comments in the operator's configuration were in Russian. One figure shows the framework inspecting an access-control check in the Spectra (Ultimate Addons for Gutenberg) WordPress plugin, version 2.19.25.
Named offensive frameworks in live intrusions. GTIG names ARTEX, CAIRN, CyberStrikeAI and PentAGI. It describes CAIRN as "an open source, multi-agent pentesting framework" in which agents share a "blackboard" and work from a starting point such as a target IP address toward an objective, and says it "observed CAIRN used in intrusion activity against targets including in the aerospace sector in the US." It says a customized PentAGI build was used between late August and early October 2026 for reconnaissance and vulnerability scanning "against a wide range of organizations in Asia." It also names ZeroHollow, an inference service used for exploit development and model routing, and Agent Zero, an open source agent framework with autonomous code execution.
Recent n-day examples. GTIG points to the PaperCut NG/MF chain (CVE-2026-81578, an authentication bypass, and CVE-2026-82078, unsafe dynamic class loading), exploited as zero-days in late August 2026, and to the PHP flaws CVE-2026-109618 and CVE-2026-109619, which it says "were discovered using LLM-enabled analysis" and where "a POC video and early patched code were published" before the official patches on October 21. All four are in CISA's Known Exploited Vulnerabilities catalog: the PaperCut pair added August 31 (due September 14), the PHP pair added October 26 (due October 29), per the CISA JSON we pulled today. CVE-2026-40217 (LiteLLM) was not in that catalog when we checked.
Credential resale. GTIG says financially motivated actors steal model credentials and cloud resources and resell access through proxy services, naming APIMart and JetNight. It also says the Chinese-language phishing service Jieyou added a chatbot in June 2026 that lets customers build phishing pages by typing plain-language instructions.
What is not known
How SOMBERMEME works. GTIG calls it a backdoor and says nothing else about it: no hashes, no command-and-control addresses, no platform. It does not say whether the settings change used a hook, which hook, or another setting. Our advice to check hooks first is our inference from how Claude Code runs commands, not something GTIG states.
How many victims. GTIG says "multiple intrusions" for MIDNIGHT NEPTUNE and gives no count of organizations, repositories or developers affected.
Who ran exploit_agent. GTIG does not attribute the staging server to a named group, does not say which vulnerabilities it successfully exploited, and does not say whether the Spectra plugin check it inspected was actually vulnerable. We found no CVE named for it in the report.
Attribution confidence. GTIG's own wording is "suspected MIDNIGHT NEPTUNE."
Indicators
GTIG published no IP addresses, domains, URLs or file hashes in this report, and links no indicator collection. The only concrete artifacts are the file path .claude/settings.json, the malware name SOMBERMEME, the framework name exploit_agent and the Copilot-themed token naming. None of those is a blockable network indicator, so we added nothing to our feed for this story. If GTIG or Mandiant publish SOMBERMEME indicators, we will add them with the source named.
How we know
We read GTIG's report directly on cloud.google.com and quoted it. We used The Hacker News only to find it. We pulled KEV dates from CISA's JSON file today. The Shai-Hulud settings-file behavior comes from our own notes on the tensorlake 0.5.144 npm package in October 2026. We tried to check our edge telemetry for probes against the Spectra plugin path, and the check could not reach our search server, so we make no claim either way about what our edge saw. We do not claim to have seen any of this activity before Google did.
Sources
Google Threat Intelligence Group, "Agentic threat activity crosses a significant threshold," November 16, 2026: cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold
The Hacker News, "Google Finds Hackers Using AI Agents to Build and Run Exploits From Software Patches," November 16, 2026: thehackernews.com/2026/11/google-finds-hackers-using-ai-agents-to.html
CISA Known Exploited Vulnerabilities catalog (JSON feed), entries for CVE-2026-81578, CVE-2026-82078, CVE-2026-109618 and CVE-2026-109619
DugganUSA, "A Worm in an AI-Agent SDK Went Looking for .claude and .cursor. Revoke the Stolen Token Too Early and It Wipes Your Home Directory," October 8, 2026
DugganUSA, "AI Agents Stole AI Keys From Exposed LiteLLM, Dify and One API Servers," November 16, 2026
How do AI models see YOUR brand?
AIPM has audited 250+ domains. 15 seconds. Free while still in beta.
Was this useful? Thirty seconds, no cookies, no tracking, no third parties, your address hashed and never stored. The question lives here: https://analytics.dugganusa.com/nps.html#post=gtig-claude-code-settings-backdoor-ai-exploit-agents



Comments