Technology

Add as a preferred source on Google

Suspected North Korean Group Turned Claude Code Settings Into Malware Route

Google also found an exposed server hosting an AI framework set up to examine software for possible flaws and automatically run drafted exploits.

Updated 2 min read
A cybersecurity analyst examines a laptop in Mountain View, California. / EconoTimes.ai (mono)
A cybersecurity analyst examines a laptop in Mountain View, California. / EconoTimes.ai (mono)

Google’s Threat Intelligence Group (GTIG) said a suspected North Korean cluster altered developer settings to deliver a backdoor into crypto firms, while a separate AI framework was configured for automated exploit work.

In a Nov. 16 quarterly update, GTIG described activity against cryptocurrency organizations in North America and Europe. It also reported finding exploit_agent on an exposed staging server.

GTIG assessed the staging scripts launched by diagnostic-fix files were likely written by large language models. The group it tracks as MIDNIGHT NEPTUNE used fabricated recruiter profiles and imitation video-call services to draw people at crypto organizations into a supposed technical-fix process. Diagnostic-fix scripts run by targets launched the staging code, GTIG said.

Repository configuration then gave attackers a route into developer tools. GTIG said operators altered .claude/settings.json; a subsequent Claude Code session fetched the SOMBERMEME backdoor. The group also tried to blend in by labeling GitHub Personal Access Tokens (PATs) as Microsoft Copilot, according to GTIG.

The separate exploit_agent framework linked local and advanced language models. Its settings called for inspecting software changes across patches, searching for possible vulnerabilities, drafting exploit code and launching it automatically. GTIG said that describes the framework’s configured capabilities, not verified activity by that instance.

The server also held exfiltrated victim data, tools for mapping internal systems and attempting brute-force access, and scripts to install or remove webshells, including those left by other attackers. GTIG said earlier work it reported had automated parts of vulnerability research; its newest finding was the first time it had identified systems configured to find weaknesses and run exploits without a person directing them.

GTIG’s tracking showed nearly twice as many vulnerability disclosures and exploits in August 2026 as in January. Zero-day exploits averaged 11 a month over the period, versus a monthly average of eight in 2025; GTIG counted 22 in August.

The campaign adds a new example to EconoTimes’ earlier coverage of AI’s reported role in North Korean targeting of crypto firms.

Related articles

View all