We manage your security to weather the AI storm →

← Resource Hub

Google: Suspected DPRK Group Abused Claude Code Settings

By Daniel GelfandPublished 5 min read

Part of our guide to onchain monitoring

Google: Suspected DPRK Group Abused Claude Code Settings

Quick answer. On November 16, 2026, Google reported that a cluster it suspects is MIDNIGHT NEPTUNE, a North Korea-nexus group, edited .claude/settings.json in crypto firms' internal repositories. Starting Claude Code downloaded the SOMBERMEME backdoor. Access began with fake recruiter profiles and ClickFix meeting lures.

Watch: Google: Suspected DPRK Group Abused Claude Code Settings 1:46
Video transcript

A suspected North Korean group edited Claude Code settings inside crypto firms' repositories, Google says. Starting Claude Code then downloaded a backdoor.

Google's November 16 report ties intrusions at crypto firms in North America and Europe to a cluster it suspects is Midnight Neptune, a North Korean group.

Fake recruiters send targets to a fraudulent meeting platform, where a ClickFix prompt asks them to run a fix script. Once inside, attackers edited the Claude settings file, so starting Claude Code downloaded Sombermeme.

Anthropic documents that file as shared project configuration, committed so one edit can reach many developers. Its hooks run commands automatically, with full user permissions.

Current docs withhold these hooks in interactive sessions until you trust the folder. Scripted and SDK sessions treat it as trusted. Google does not say which applied here.

Google names no victims, stolen amounts or remediation status, and reports no Claude Code vulnerability. As of November 18, we found no recovered funds or regulatory steps.

Treat assistant settings as code. Search repository history, require code-owner review and alert on hook changes. Enforce managed hook policy, review GitHub tokens by creator, not name, and refuse diagnostic scripts from meeting pages.

In Guardian's analysis, development access should never imply signing authority. Read the full article, or talk to Guardian about an OpSec Audit of access, secrets and signing.

What Google Disclosed

On November 16, 2026, Google Threat Intelligence Group published an AI threat tracker report on agentic attacker activity. BleepingComputer reported it the same day.[1, 2]

One case concerns cryptocurrency organizations in North America and Europe. Google attributes multiple intrusions there to a suspected MIDNIGHT NEPTUNE cluster, which it places among North Korean actors.

The report names the file they changed. After getting in, they edited .claude/settings.json files inside the victims' internal code repositories.

Google says that when users then started Claude Code, it triggered a download of the SOMBERMEME backdoor.

Google places the intrusions in 2026 but gives no precise dates. This is a disclosure of past activity, not a report of a new intrusion this week.

The Attack Chain

Google describes a chain that starts with people and ends inside developer tooling.[1]

StageWhat Google reports
ApproachRecruiter profiles with SynthID-watermarked images, which Google says confirm LLM use
MeetingTargets sent to fraudulent meeting platforms
LureClickFix prompt to run a diagnostic fix script
StagingLikely LLM-authored staging scripts
PivotEdited .claude/settings.json in internal repositories
CoverGitHub tokens named after Microsoft Copilot, masking lateral movement

Why a Settings File Can Run Code

Anthropic documents .claude/settings.json as shared project configuration. It can hold permissions, hooks, plugins and environment values.[3]

Teams commit it so teammates who clone the repository get the same settings. One edit can therefore reach many developers.

Hooks are commands that run automatically at points in a session. Anthropic warns that command hooks execute with the user's full permissions.[4]

In interactive sessions, current documentation withholds hooks from project settings files until the user trusts that folder or a parent folder.[5]

Scripted -p and SDK sessions show no dialog and count the folder as trusted. Settings-file hooks then run even in a folder never trusted before.

Google does not name the settings key the attackers used, the client versions, or whether victims had already trusted those repositories.[1]

Precedence shapes defense. Managed settings generally outrank project settings, and shared project settings outrank a developer's own user file.

A user-level disableAllHooks setting can be overridden by project settings. Managed allowManagedHooksOnly blocks user, project, local and plugin hooks, except plugins force-enabled by policy.

What Is Still Unknown

Google names no victims and gives no victim count, stolen amounts or remediation status. As of November 18, we found no later disclosures, recovered funds or regulatory steps.[1]

Google calls SOMBERMEME a backdoor but does not describe its commands, supported platforms or exfiltration functions.

Its general mitigation passage says Google disabled assets and updated its models. That passage covers the whole report, not the crypto intrusions specifically.

BleepingComputer calls the cluster UNC6836, with suspected links to Midnight Neptune. That label mapping is the publisher's wording, and attribution remains an intelligence assessment.[2]

BleepingComputer also relays Google researchers' outlook: unsophisticated, mostly multi-agent attacks within three months, and sophisticated, fully autonomous ones within six. Those are forecasts, not a finding that these intrusions ran autonomously.

What Is New Here

North Korean lures against crypto staff are not new. Our FaceStealer explainer covers the FBI's November 12 disruption of a fake video-call scheme.[6]

Google's September tracker says MIDNIGHT NEPTUNE was formerly tracked as UNC1069, the label Google gave the March 2026 axios compromise.[7, 8]

Our axios npm attack explainer covers that fake-call case. A shared label does not show that these intrusions and that case were one operation.

Our guide to North Korea hiring scams covers fake recruiters whose coding tests plant BeaverTail and InvisibleFerret. Google's report does not connect MIDNIGHT NEPTUNE to those campaigns.[1, 9]

Google's tracker dated September 8 already listed poisoned repository configurations, altered Claude CLI hooks and SOMBERMEME deployment. The November report adds the .claude/settings.json path and the startup download.

The foothold is the difference from earlier lures. The attacker sits in a file the team's own assistant reads, inside the team's own repository.

Our dApp frontend supply chain security guide covers npm packages and install scripts. A tampered settings file is not a dependency, so lockfile and package controls may never see it.[10]

What to Check Now

Treat assistant configuration as code that executes, because it can.

Anthropic documents a ConfigChange hook that runs on direct edits to settings files. Set centrally, it can log those edits for review.[3]

It fires on changes during a running session, so it will not flag a tampered file already in place when Claude Code starts. Treat it as a supplement to commit review.[4]

Plugins raise the same trust question. Our Trail of Bits skills guide sets out how to vet a plugin's hooks and scripts before enabling it.

Device, account and recruiter rules are in our crypto OpSec baseline.

  • Search every repository's history for 2026 changes to .claude/settings.json and other assistant configuration files.[1]
  • Require code-owner review for .claude/ directories and similar assistant configuration paths.
  • Alert on any commit that adds or changes hooks, plugins or environment values.
  • Where teams do not need project hooks, set allowManagedHooksOnly in managed settings.
  • For scripted runs, pass --setting-sources user on the CLI, or set the SDK's settingSources without project, unless the job needs project settings.[5]
  • Review GitHub personal access tokens by creator, scope and last use, never by name alone.
  • Revoke tokens nobody can account for, then rotate the secrets they could reach.
  • Refuse any diagnostic script offered by a meeting page or a recruiter.
  • Open unfamiliar repositories with an assistant only in environments holding no production credentials.

Guardian Analysis: Separate Code From Keys

This section is Guardian analysis. Google's chain runs through a developer's account, so what that account can reach sets the damage.

A backdoor launched by a developer's assistant runs as that developer. Key material left by ordinary engineering workflows is then within reach.

Guardian's public Bounce Tech report records finding I-05, rated Informational. A shell script accepted a private key as a command-line argument.[11]

The finding notes the key could surface in shell history, process listings and logs, where anyone with the user's account, process access or those logs could retrieve it. The report marks it resolved.

I-05 illustrates key handling risk. It is not the configuration tampering Google describes, and not evidence that this campaign touched Bounce.

The practical lesson is separation: development access should never imply signing authority.

Frequently Asked Questions

Does this mean Claude Code has a vulnerability?

Google reports no Claude Code vulnerability. We found no Anthropic statement on this campaign, and the documentation cited here is not a fix or response to it.[1]

Are other coding assistants involved?

Google's report names only Claude Code settings files. Any assistant that reads repository configuration and can run commands deserves the same review, but no abuse of other products is reported here.[1]

What if we find an unexpected settings change?

Treat every machine that started Claude Code in that repository since the change as suspect. Isolate those machines, then identify which account and token pushed the edit before trusting the repository again.

Sources

  1. Google Threat Intelligence Group: GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold
  2. BleepingComputer: Google warns that fully-automated AI attacks are coming soon
  3. Anthropic: Settings files and precedence
  4. Anthropic: Hooks reference
  5. Anthropic: Configure permissions
  6. Guardian: FBI Disrupts FaceStealer: North Korea's Fake Call Scheme
  7. Google Threat Intelligence Group: GTIG AI Threat Tracker: From Prompting to Autonomy - The Evolution of Adversarial AI
  8. Google Threat Intelligence Group: North Korea-Nexus Threat Actor Compromises Widely Used Axios NPM Package in Supply Chain Attack
  9. Guardian: North Korea Hiring Scams: Fake Interviews and Coding Tests
  10. Guardian: dApp Frontend Supply Chain Security: npm Attack Mitigation
  11. Guardian: Bounce Tech: Automation Layer and Contract Updates
  12. Guardian: 246 Club: re-a-token
  13. Guardian: OpSec Audit for Digital Asset Teams

Review what a compromised developer can reach

Guardian's OpSec Audit reviews access and change control, secrets management, environment separation and signing procedures, the boundaries a compromised developer account can cross.