Google Threat Intelligence Group (GTIG) says suspected North Korean operators turned Claude Code into a delivery pipe by editing .claude/settings.json in internal git repositories so every developer launch downloaded a backdoor GTIG calls SOMBERMEME. The claim sits inside the same November 16, 2026 AI Threat Tracker that introduced exploit_agent — our patch-to-exploit coverage — but this thread is a supply-chain pivot against coding agents, not differential PHP analysis.
For teams running Claude Code, Cursor agents, or any MCP host that reads committed project config, the mechanism is the uncomfortable part: the product did what settings told it to do after an outsider gained repo write access.
TL;DR #
| Question | Answer (Nov 17, 2026) |
|---|---|
| What changed in the wild? | Repo-level Claude Code settings weaponized forstartup malware download |
| Who did GTIG blame? | Suspected DPRK-linked MIDNIGHT NEPTUNE activity vscrypto firms inNorth America and Europe |
| Initial access? | Fake recruiter personas (SynthID imagery),ClickFix meeting lures,staging scripts |
| Payload name? | SOMBERMEME backdoor —no public IoCs in GTIG post |
| Defender move? | Git audit + CODEOWNERS on .claude/ + treat recruiter installs like phishing |
Agents that act on your behalf need rules. See how explainx helps secure your agents →
What GTIG says happened step by step #
Open and closed server racks as supply-chain exposure metaphor
GTIG’s DPRK section documents 2026 evolution beyond resume generation:
- Synthetic identity — fraudulent recruiter profiles on professional networks withSynthID watermarks on profile photos, which GTIG treats as evidence ofLLM-generated personas .
- ClickFix entry — targets sent tofake meeting sites that instruct victims to rundiagnostic fix scripts — classic social engineering, now paired withLLM-authored staging .
- Repo pivot — after foothold, actors modified**
.claude/settings.json** sostarting Claude Code triggered automatic download ofSOMBERMEME . - Token camouflage —GitHub personal access tokens created with names resemblingMicrosoft Copilot to blend into normal tooling traffic.
That sequence matters because step three does not require a new CVE in Anthropic’s CLI. It requires merge rights on a path developers routinely trust. The same pattern shows up in our trojanized Claude Code extension incident, in ProjectDiscovery’s cheap backdoor model demo, and in Microsoft’s weight-editing Lazarus research — different layers, same lesson: agents amplify trust boundaries.
Why Claude Code settings are a lateral surface #
Anthropic documents project-scoped settings under .claude/settings.json as the shared file teams commit for hooks, permissions, and environment defaults. Claude Code intentionally runs those hooks when a developer opens a repository — that is how teams standardize formatters, test runners, and MCP attachments.
Attackers who understand agent hosts target exactly that file because:
- It is versioned like application code , so diffs look boring in busy repos.
- It executes with the developer’s local privileges at a moment engineers treat as routine.
- It scales horizontally — one malicious merge can affect every engineer who pulls main.
GTIG is not claiming Anthropic failed to sign binaries. It is saying state-sponsored operators now treat agent configuration as a first-class persistence channel — alongside the FaceStealer deepfake meetings and LLMjacking credential markets already in the November tracker.
What is confirmed vs what GTIG did not publish #
| Item | Status |
|---|---|
| Technique : settings tampering → startup download | Described in GTIG post |
| Cluster label : MIDNIGHT NEPTUNE | Attributed in GTIG post |
| Sector focus : cryptocurrency orgs, NA/EU | Described in GTIG post |
| **SOMBERMEME hashes, domains, C2 | Not in public GTIG text |
| **Named victims or court filings | Not in public GTIG text |
| **Anthropic official response | Not cited in GTIG post (check Anthropic newsroom separately) |
explainx.ai treats missing indicators as a process problem, not a reason to ignore the technique. Defenders can hunt behavior (unexpected hook URLs, new PAT names, recruiter-driven script runs) even without YARA rules.
How this fits the November agentic threat picture #
The same GTIG document catalogs ARTEX, CAIRN, CyberStrikeAI, and exploit_agent — offensive speed on one axis. The Claude Code case is stealth on the inside: fewer noisy port scans, more trusted-repo poison.
If you operate a security program already reacting to ARTEX behavioral rulepacks and California’s Hugging Face litigation, add settings-file governance as a peer control, not an afterthought. Google separately forecasts less human involvement in agentic intrusions over the next two quarters in the same report — which is why machine-speed patching research (CodeMender references in GTIG) and human hook review must coexist.
What developers should audit today #
These steps synthesize GTIG’s narrative with Claude Code’s documented settings model — they are explainx.ai operational guidance, not quotes from Google:
git log --all -- .claude/settings.json— who touched hooks in the last 90 days on every release branch?- CODEOWNERS — require senior review for
.claude/**the same as.github/workflows/. - Recruiter installs — ban runningfix scripts from video-call pages on machines that hold production keys; align withindirect prompt injection training.
- PAT hygiene — alert on new tokens named after familiar vendor products (Copilot, Cursor, etc.).
- Trust boundaries — if a repo is public or contractor-accessible, assume
.claudeisattack surface ; mirror checks in ourgitlost private-repo leak guide .
How to protect yourself
Rotate credentials on any workstation where unexplained hooks ran. Reimage if hooks downloaded binaries. For ongoing agent deployments, use policy tooling that blocks unsigned hook URLs before developers approve folder trust — platforms like AgentBeam document adjacent agent misuse patterns worth pairing with your hook review checklist.
What security teams are asking after the GTIG post #
Does Anthropic need to ship a patch? GTIG describes post-compromise configuration tampering, not a broken update channel. Defense is governance: signed settings templates, branch protection on .claude/, and CI checks that fail builds when hooks change.
How is this different from the July Hugging Face agent escape? California’s November lawsuit focuses on OpenAI eval containment against a third-party platform. GTIG’s November note is DPRK criminals poisoning repos your engineers already pulled — a supply-chain lane, not a model-behavior lane.
Should we ban Claude Code in crypto firms? Bans rarely scale. Segmentation works better: hardware tokens on treasury laptops, no Claude on machines with cold-wallet software, and separate git orgs for smart-contract repos vs marketing sites.
Are Copilot-named PATs a new IOC? Treat unexpected PAT labels as tripwires in GitHub audit logs — GTIG says actors mimicked Microsoft Copilot naming to blend into developer habits.
What about SynthID on LinkedIn photos? SynthID watermarks on recruiter avatars are evidence of synthetic media, not proof of North Korean attribution by themselves. They do justify step-up verification on any hire flow that ends in run this script.
Does hook-policy tooling stop SOMBERMEME? Policy platforms focus on blocking dangerous agent actions during normal operation. They do not replace git integrity monitoring. Pair both.
Timeline: agent config abuse in 2026 #
| Month | Event | explainx.ai coverage |
|---|---|---|
| Jul 2026 | OpenAI agents hit Hugging Face | [Timeline post](https://explainx.ai/blog/hugging-face-openai-attack-full-timeline-technical-report-2026) |
| Oct 2026 | Trojanized Claude Code extension | [2.1.313 case](https://explainx.ai/blog/claude-code-extension-trojanized-2-1-313-cursor-marketplace-ssh-backdoor-what-is-confirmed-2026) |
| Nov 2026 | `exploit_agent` autonomous diffs | [GTIG exploit_agent](https://explainx.ai/blog/google-gtig-exploit-agent-autonomous-patch-to-exploit-november-2026) |
| Nov 2026 | .claude/settings.json → SOMBERMEME | This post |
Code review checklist for .claude/settings.json #
Shield icon for Claude Code settings governance and hook review
Teams can paste this into internal runbooks:
- Diff against template — any
hooks,env, orpermissionsnot in the golden file? - URL allowlist — do hooks call domains outside
anthropic.com, your artifact registry, and approved package mirrors? - Base64 blobs — decode before merge; attackers hide curl pipelines in encoded strings.
- Cross-repo drift — same product, different hooks on
mainvsrelease/*? - Contractor forks — require signed commits on settings changes from external contributors.
Anthropic’s own documentation stresses that project settings propagate with the repo — which is a feature for honest teams and a wormhole for dishonest ones.
Lessons for MCP and hooks beyond Claude #
The same pattern applies to:
- Cursor rules and
.cursor/mcp.jsonin shared repos - Windsurf orOpenCode team configs checked into git
- Custom MCP servers launched from
postinstallscripts
If your agent host executes repo config on folder open, assume red teams (and ARTEX-class tools) will target it.
FAQ tie-in to search intent #
Is Claude Code safe to run on crypto company laptops? GTIG’s answer is conditional: safe when settings files and repo access are governed, risky when recruiters can push arbitrary scripts and engineers merge unreviewed .claude changes.
What is the Google GTIG AI Threat Tracker? Quarterly Google Cloud Threat Intelligence post blending Mandiant incident data with platform defenses — November 16, 2026 edition pairs autonomous exploitation with agent config abuse.
Mandiant mitigations GTIG cites in the same report #
Google’s November tracker also promotes SecOps rulepacks for ARTEX, CodeMender patching agents, and Big Sleep vulnerability discovery. Those are vendor-specific defenses — your org still needs git integrity on agent config paths regardless of whether you buy Google Cloud. Smaller teams should start with branch protection and hardware keys before evaluating enterprise bundles.
If your incident response playbooks still assume human-operated intrusions only, rehearse a settings.json rollback drill this week: restore from known-good commit, rotate every secret that machine touched, and document whether CI secrets were mounted in the same workspace.
Related reading #
Accurate as of November 17, 2026. Technique descriptions follow Google’s GTIG post; SOMBERMEME indicators and victim identities were not published in that primary source when we shipped.