North Korean hackers turn Claude Code into a delivery channel for a backdoor
According to the Google Threat Intelligence Group, suspected North Korean attackers tampered with Claude Code's configuration file in company repositories. From then on, every launch of the AI coding tool automatically downloaded a backdoor called SOMBERMEME. They didn't need a vulnerability at Anthropic. All they needed was write access to a file developers trust blindly.

What's going on
The Google Threat Intelligence Group (GTIG) is Google's threat analysis unit, which also includes the security firm Mandiant. On November 16, 2026, it published the latest edition of its "AI Threat Tracker", a regular overview of how attackers use AI. One case in it hits developers who code with AI agents directly: suspected North Korean attackers allegedly used Claude Code, Anthropic's command-line coding tool, to distribute malware.
To do this, they modified the .claude/settings.json file in internal Git repositories. Git is the standard system teams use to manage code together. After the change, every launch of Claude Code automatically downloaded a backdoor that Google calls SOMBERMEME. A backdoor is malware that lets attackers control a computer remotely. This account is based on an analysis by the blog explainx.ai. We have not seen the original Google report ourselves.
How the attack worked
According to GTIG, the campaign ran in four stages:
- Fake identities: The attackers posed as recruiters on career networks. The profile photos carried SynthID watermarks, an invisible marker Google embeds in AI-generated images. GTIG takes this as a sign that the personas were created with AI.
- ClickFix as the entry point: Targets were lured to fake meeting pages, where they were told to run a supposed diagnostic or repair script to "fix" a technical problem. This technique is called ClickFix. It's old, but according to GTIG, the scripts for the next stage now come from language models.
- Jumping into the repository: With the access they'd gained, the attackers changed the Claude Code settings in the company repository so that SOMBERMEME was downloaded at startup.
- Camouflage: To access GitHub, they created Personal Access Tokens. Their names resembled Microsoft Copilot so they wouldn't stand out among normal developer tools.
Google attributes the activity to a cluster called MIDNIGHT NEPTUNE. The targets were cryptocurrency companies in North America and Europe. That fits earlier reports that North Korean groups mainly go after crypto firms to make money.
Why the settings file is so dangerous
Here's the uncomfortable part: Claude Code had no bug. The program did exactly what its configuration told it to. Anthropic provides a shared .claude/settings.json file for projects, which teams check into the repository alongside their code. Among other things, it contains hooks, commands that run automatically on certain events. It also holds permissions, environment variables and connections to MCP servers. MCP (Model Context Protocol) is a standard that lets AI agents access external tools and data sources. Teams use all this for perfectly legitimate reasons, for example to configure formatters or test runs the same way for everyone.
According to explainx.ai, this file makes a good attack vector for three reasons:
- It's versioned like regular code. In busy repositories, small changes to it easily slip through.
- It runs with the developer's local privileges, and at a moment nobody considers risky: when opening the project.
- It affects everyone. A single malicious change on the main branch reaches every developer who pulls the latest version.
So GTIG isn't claiming that Anthropic, say, signed binaries sloppily. The point is this: state-backed attackers are now deliberately using AI agent configuration to entrench themselves in systems for the long term. Security people call this persistence. explainx.ai sums it up like this: project settings travel with the repository. For honest teams, that's a useful feature. For dishonest ones, it's a "wormhole".
What's confirmed and what's missing
| Item | Status |
|---|---|
| Technique: tampered settings trigger a download at startup | described in the GTIG report |
| Attribution to MIDNIGHT NEPTUNE | named in the GTIG report |
| Targets: crypto firms in North America and Europe | described in the GTIG report |
| Hashes, domains, C2 servers for SOMBERMEME | not published |
| Victim names, court records | not published |
| Official statement from Anthropic | not quoted in the GTIG report |
That means defenders don't have ready-made indicators like file checksums or antivirus signatures. explainx.ai therefore recommends hunting for behavioral patterns: hook commands that call unexpected addresses, new tokens named after products, and scripts run after contact with a "recruiter".
The same Google report describes other developments too, including exploit_agent, a tool that automatically derives attacks from security patches, as well as tools like ARTEX, CAIRN and CyberStrikeAI. Google expects fewer and fewer humans to be involved in agentic attacks over the next two quarters. The Claude Code case is the flip side of those loud, fast attacks: little network noise, just one poisoned file in a trusted repository.
What teams should check now
The following recommendations come from explainx.ai, not Google. They combine GTIG's description with the official Claude Code documentation:
- Review the change history: Run
git log --all -- .claude/settings.jsonto see who changed the file on which branch over the last 90 days. - Require reviews: Use a CODEOWNERS file to require that changes to
.claude/**be approved by senior people, just as is already common for CI workflows under.github/workflows/. - Compare against a template: Any deviation from a vetted baseline configuration in hooks, environment variables or permissions is suspicious.
- Check addresses: Hooks should only reach approved destinations, such as your own package registry. Decode Base64-encoded blobs before merging, because that's a favorite hiding place for download commands.
- Token hygiene: New GitHub tokens with names like "Copilot" or "Cursor" should trigger an alert.
- Treat recruiter scripts like phishing: Nobody should run scripts from video call pages on machines that hold production keys.
If you find unexplained hooks on a machine, explainx.ai says you should rotate all credentials on it. If a program was downloaded, the machine should be wiped and rebuilt from scratch. For crypto firms, the blog recommends separation over bans: no Claude installs on machines running cold wallet software, hardware security keys for treasury laptops, and separate Git organizations for smart contract code.
The problem isn't limited to Claude Code. Any agent that executes configuration from the repository when you open a folder is vulnerable in the same way. That includes .cursor/mcp.json in Cursor, team configurations in Windsurf or OpenCode, and MCP servers launched via install scripts.
The bigger picture
The case shows how attacks are shifting as AI agents move into developers' daily work. The attackers don't crack the tool. They exploit the trust teams place in it and its configuration. A patch from Anthropic wouldn't change much, since the behavior described works as designed. What's needed is process: branch protection, mandatory reviews and automated checks that fire whenever hooks change. Some protection comes from Claude Code asking whether you trust a folder the first time you open it. But for a repository you use every day, that prompt barely helps.
A caveat: our source is a blog that summarizes the Google report and adds its own advice plus promotion for agent security tools. Technical evidence like malware checksums or server addresses isn't publicly available, and Anthropic hasn't commented so far. It's unclear how many companies are affected. And the SynthID watermarks on their own only prove that the images were AI-generated. They don't prove North Korea is behind it. Still, the lesson applies to every team using AI agents: agent configuration files are executable code and deserve the same scrutiny.
Sources
- Merkliste: explainx.ai explainx.ai
Ask the AI assistant
Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.
The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.
Comments
What do you think? Already tried it? Write it here, no sign-up needed.
Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.