Google Threat Intelligence Group (GTIG) says a suspected North Korean cluster it tracks as MIDNIGHT NEPTUNE altered committed .claude/settings.json files so everyday Claude Code launches downloaded the SOMBERMEME backdoor—after fake recruiter personas and ClickFix lures got humans to run staging scripts. The Q4 2026 AI Threat Tracker, published November 16, 2026, places the technique under "Targeted Espionage and Supply Chain Evolution" and ties it to intrusions at cryptocurrency firms in North America and Europe.
This is not a remote Claude Code zero-day in GTIG's telling. It is configuration supply chain: attackers with repo access turned trusted agent hooks into a launch-time downloader.
| Question | Short answer |
|---|---|
| Am I affected? | Only if a machine or repo you use was compromised and settings files were tampered with—GTIG names crypto sector victims, not all Claude users |
| Is there a fix? | No vendor patch for "malicious hook" content—revert unauthorized settings, reimage hosts, rotate secrets |
| Does data leave my machine? | SOMBERMEME is described as downloaded at launch—treat affected PCs as fully compromised |
| Are IOCs public? | No hashes/domains in the GTIG blog section Beam cited |
| What should I audit today? | Git history of .claude/settings.json, local and managed settings, PAT names mimicking Copilot |
explainx.ai summarized the same GTIG release alongside exploit_agent reporting for readers following AI threat news timelines.
Attack chain step by step (per GTIG)
GTIG's public narrative, quoted and paraphrased from the November 16 tracker:
- Synthetic recruiter persona. Profiles on professional networking sites used SynthID watermarks on profile imagery—GTIG treats that as confirmation of LLM-generated deceptive identities.
- ClickFix lure. Targets reached fraudulent meeting platforms that asked them to run "diagnostic fix" scripts—staging access on the workstation.
- Repository pivot. With foothold, the actor edited internal code repositories.
- Settings supply chain. Modifications to
.claude/settings.jsonensured starting Claude Code automatically triggered download of SOMBERMEME. - Stealth tokens. GTIG adds that the actor created GitHub personal access tokens named after Microsoft Copilot to blend lateral movement with routine tool usage.
GTIG also notes Google disabled assets tied to misuse and updated safeguards, in the same report's mitigation section—without publishing victim-specific indicators for SOMBERMEME itself.
Why Claude Code settings are a high-value target
Anthropic documents project-scoped settings under .claude/settings.json as committable configuration—hooks, permissions, and environment defaults flow through Git like CI YAML. Beam's agent hooks documentation lists ~/.claude/settings.json and PreToolUse as the Claude Code integration surface.
That design is correct for teams who want shared deny rules and consistent hooks. It also means:
- A malicious commit to
.claude/settings.jsonpropagates the same way a maliciouspackage.jsonscript would. - SessionStart, PreToolUse, and related hook slots run with developer trust—users who already approved the folder may not re-read every hook line each launch.
- Attackers do not need a marketplace extension compromise when they can edit internal monorepo config—contrasted with the compromised Claude Code VS Code 2.1.313 extension case, which abused marketplace trust instead.
GTIG's Copilot-named PAT detail is a social layer on top: even token inventory dashboards may dismiss those credentials as benign Microsoft tooling until someone reads scopes.
What GTIG did not publish (and how to defend anyway)
The Q4 tracker paragraph Beam relied on does not include:
- File hashes, C2 domains, or malware family teardown for SOMBERMEME.
- A fixed Claude Code version—because the abuse path is configuration, not a cited CVE.
Defensive work therefore looks like supply-chain hygiene, not claude --version alone:
# Paths GTIG and Anthropic docs treat as hook-bearing (review contents and Git history)
git log --oneline -- .claude/settings.json .claude/settings.local.json
cat .claude/settings.json
cat ~/.claude/settings.json 2>/dev/null || true
Extend review to any hook command that shells out to curl, wget, powershell, bash -c, or unknown package installs. Compare against your managed settings policy if your org uses Anthropic enterprise controls.
Pair settings review with human-factor controls from the same GTIG report: recruiter lures and ClickFix are upstream of the settings edit—training and application allowlisting matter before Claude Code starts.
What this means if you run AI coding agents
- Treat
.claude/like.github/workflows/. Require code review, CODEOWNERS, and CI grep for hook blocks on every merge—especially monorepos crypto or fintech teams GTIG names. - Baseline hooks per repository. Store known-good
settings.jsonfingerprints; alert on drift outside approved change windows. - Do not trust PAT display names. Audit GitHub PAT scopes and creation times; GTIG explicitly calls out Copilot-mimicking names.
- Reimage after suspected hook tampering. A launch-time downloader compromises the whole interactive session, not only the Git tree.
- Cross-read parallel DPRK campaigns. The DOJ FaceStealer disruption shows the same actor economy targeting developers through meetings and bogus scripts—SOMBERMEME is the agent-config variant once they are inside.
Connect to broader agentic intrusion reporting: Cisco Talos on ARTEX-class tooling and GTIG exploit_agent patch weaponization describe speed on the offensive side; SOMBERMEME describes persistence on the developer desktop through a tool developers open daily.
Same GTIG report, different failure modes
The Q4 AI Threat Tracker is not only about DPRK settings tampering. Published the same day, it documents exploit_agent staging servers that diff patches and auto-fire exploits, LLMjacking bills above $100,000 in a single day, and ARTEX-class frameworks in live Mandiant cases. Those threads matter for coding-agent users even when SOMBERMEME never touches their laptop:
| GTIG theme | Risk to coding-agent users | Beam-relevant response |
|---|---|---|
| exploit_agent | Shorter safe windows for web apps agents deploy | Patch what agents ship; deny destructive deploy tools locally |
| LLMJacking | Stolen API keys resold via mixers | Scope keys; watch anomalous inference in cloud accounts |
| MIDNIGHT NEPTUNE / SOMBERMEME | Committed hook config on dev machines | Git-review .claude/settings.json; reimage on tampering |
SOMBERMEME sits at the intersection of North Korean IT-worker tradecraft (fake recruiters, scripted meetings) and native agent hooks—the same report family that notes DPRK actors using LLMs for résumés and interview prep. After foothold, weaponizing Claude Code's intended configuration channel is a logical next step: developers already trust those files to install team hooks and allow rules.
BleepingComputer's November 16 coverage of the tracker repeats the settings.json backdoor paragraph for readers who do not read the full Google Cloud tracker on first open—useful as secondary confirmation of the SOMBERMEME chain, not a replacement for the primary GTIG text.
Managed settings and enterprise Claude Code
Anthropic's enterprise path can centralize restrictions on which settings files apply and what hooks are permitted; exact knobs evolve by release. GTIG's scenario assumes attackers could commit malicious project settings into repositories developers already clone. Mitigations security teams discuss in parallel with this incident:
- CODEOWNERS on
.claude/and rejection of hook changes without security review. - CI checks that fail builds when
settings.jsoncontains unexpectedhooksblocks or outbound download commands. - Separate admin workstations for treasury or key-management tasks, so a recruiter lure on a marketing laptop never shares disk with production signing keys.
- Managed device policies that block execution of scripts from ClickFix-style meeting pages—upstream of any agent launch.
None of that is automatic in a default solo-developer install. That gap is why GTIG classifies the vector as supply chain evolution: the product feature (shared hooks) is working as designed while the organization's change control failed.
Distinguish SOMBERMEME from Microsoft "Lazarus" model backdoors
Beam covered Microsoft Research's Lazarus weight-editing backdoor in November 2026—a tensor poisoning technique for open-weight models, not the DPRK MIDNIGHT NEPTUNE name GTIG uses for intrusion clusters. The similar "Lazarus" wording is coincidence in naming layers. SOMBERMEME is host malware delivered at Claude Code launch, not a dormant neuron in a checkpoint. Your checks are Git, EDR, and settings diff, not beam skills scan on a Hugging Face download—though teams running local models behind agents should keep both supply chains in separate runbooks (open-weight sleeper discussion).
Incident response sequence if you suspect hook tampering
GTIG's public write-up does not include a vendor IR phone number for SOMBERMEME; response follows standard compromise playbooks with agent-specific extras:
- Stop launching Claude Code on affected machines until settings are verified from a known-good commit.
- Export and diff all hook-bearing JSON under
.claude/and~/.claude/against last week's baseline. - Revoke and reissue GitHub PATs created during the incident window—including any whose display name mimics Copilot, per GTIG's stealth note.
- Notify crypto or treasury teams if the developer laptop had wallet browser profiles or signing tools; MIDNIGHT NEPTUNE's published targeting emphasizes cryptocurrency firms.
- Preserve logs from EDR and any local agent collectors (including Beam) for the period before reimage, if your policy allows on-device retention.
Report new IOCs through your national CERT if you obtain malware samples beyond what GTIG published.
How Beam handles this
Beam installs PreToolUse observation and enforce hooks in Claude Code's existing settings.json merge path (agent hooks docs). It preserves other hooks rather than replacing the file wholesale—useful for visibility, not a guarantee nobody else edited the same JSON on another branch.
Beam's skill and MCP scanner (skill-testing cookbook) targets skills and MCP configs, not every key inside settings.json. Unexpected SessionStart downloaders may still run before your next hook fires unless you catch them in Git review or managed settings.
Where Beam helps after hooks are clean:
- Observe / enforce on risky tool proposals (destructive shell, credential paths) per workspace policy.
- Local-only event storage with redaction—contrasts with cloud SIEM models when assessing post-intrusion agent behavior.
Limits to state plainly: Beam is heuristic, hook-based, and fail-open on some compatibility edges documented in policies—it is not a replacement for reimaging a workstation GTIG would classify as SOMBERMEME-positive.
Frequently asked questions
What is the SOMBERMEME backdoor in Google GTIG reporting?
In the November 16, 2026 Q4 AI Threat Tracker, GTIG describes SOMBERMEME as malware delivered after MIDNIGHT NEPTUNE edited .claude/settings.json so Claude Code downloaded it at startup. GTIG did not publish file hashes in the blog post.
Is SOMBERMEME a vulnerability in Claude Code itself?
GTIG frames it as post-compromise abuse of legitimate configuration and hooks after social engineering, not a cited CVE in Anthropic's binary.
Which organizations did GTIG say were targeted?
GTIG cites cryptocurrency organizations in North America and Europe without naming individual victims in the public section.
How did MIDNIGHT NEPTUNE get initial access?
GTIG says fake recruiter profiles and ClickFix lures led victims to run staging scripts from fraudulent meeting sites.
What should I check in Claude Code settings files?
Review project .claude/settings.json, .claude/settings.local.json, and ~/.claude/settings.json for hooks or environment values you did not add.
Did Google publish SOMBERMEME IOCs?
The Q4 tracker narrative does not include hashes, domains, or YARA rules for SOMBERMEME in the MIDNIGHT NEPTUNE section.
How does this relate to DPRK open-source targeting?
It parallels maintainer-focused campaigns such as FaceStealer, but pivots into agent hook configuration on workstations after access.
Can Beam detect malicious Claude Code settings hooks?
Beam scans skills and MCP configs and watches PreToolUse; it does not replace Git review of committed settings. Treat hook edits as supply-chain changes.
Related reading
-
Darktrace: Claude memory stores abused as cloud C2 — Signal Labs showed stolen Console credentials can queue jobs through api.anthropic.com without editing repo settings.
-
Phantom53: Claude Code artifacts in malware source — SOCRadar found
.claude/settings.jsonand 143 task files inside a ClickFix implant repo, separate from victim-side settings backdoors.
GTIG quotations and incident characterizations reflect Google's November 16, 2026 AI Threat Tracker and related Google Cloud threat intelligence posts as of publication.

