Suspected North Korea group turned Claude Code into a SOMBERMEME backdoor vector, Google says

Google Threat Intelligence Group (GTIG) has revealed that a suspected North Korea-linked cluster tracked as MIDNIGHT NEPTUNE turned Claude Code, Anthropic's coding assistant, into an infection vector: in intrusions against cryptocurrency companies in North America and Europe, the attackers edited .claude/settings.json files in internal repositories so that launching Claude Code would automatically download the SOMBERMEME backdoor. The case was documented in Google's quarterly AI threat report, "Agentic Threat Activity Crosses a Significant Threshold", published on November 16, 2026.

According to the report, initial access began with fake recruiter profiles on professional networking platforms, using AI-generated photos — Google confirmed the use of generative models through SynthID watermarks in the images. Victims were directed to fake meeting platforms with ClickFix lures that prompted them to run "diagnostic" scripts. After the initial compromise, the group moved into the companies' internal code repositories and tampered with the Claude Code configuration.

The core of the attack is that Claude Code itself did not need to be compromised: the trusted tool read a tampered project configuration and executed what was in it. To blend in with legitimate traffic during lateral movement, the attackers also created GitHub personal access tokens named after Microsoft Copilot, according to Google.

What Google did

Google said it disabled the assets associated with the activity and that Google DeepMind used the findings to strengthen classifiers and models, so they refuse to assist with this kind of attack. The public text of the report does not include technical indicators for the backdoor, such as hashes or domains.

Why it matters

The case shows a new supply-chain attack frontier: instead of exploiting a flaw in the software, the attacker poisons the configuration that a trusted AI tool reads at startup. Teams that use coding assistants with hooks and shared configuration files in repositories need to audit those configurations as they would any executable code — and review who can change them.

The same GTIG report also documents the exploit_agent framework, which turns patches into exploits without human intervention — a topic Geekiki already covered in our article on exploit_agent.

Sources

Image: Anthropic logo, via Wikimedia Commons.

Transparency note: this article was produced with the assistance of artificial intelligence, based on cross-checking public posts on X and sources on the internet. The facts were verified against the original sources cited; please consult them to confirm the information.

By Muse