AI Aulis

AI NEWS IN ENGLISH · 18 November 2026

The most important AI news every morning. AI writes and checks the stories against the original sources – check important facts in the sources.

32 stories · next issue tomorrow morning · How the paper is made →

← Issue 18 November 2026Paper

Google: attackers spread a backdoor through breached companies’ Claude Code settings

AI-generated contentWritten by AI

cloud.google.com, advisories.checkpoint.com, helpnetsecurity.com, threatfrontier.com · 4 sources · Sources published: 16–17 November 2026 · Some source publication dates are unknown

Attackers changed .claude/settings.json files in the internal repositories of cryptocurrency companies they had breached so that Claude Code loaded the SOMBERMEME backdoor at start-up, Google’s threat intelligence group said.

Google Threat Intelligence Group (GTIG) describes breaches at cryptocurrency companies in North America and Europe in its AI Threat Tracker report, published this week. GTIG suspects that a North Korea-linked group it tracks as MIDNIGHT NEPTUNE was behind them. The report marks the actor as suspected, not confirmed.

The attackers set up fake recruiter profiles on professional networking services. The profile photos carried SynthID watermarks, which GTIG said show the images were AI-generated. The recruiters directed targets to fake meeting platforms, where a ClickFix lure asked them to run a ‘diagnostic fix script’.

After the breach, the attackers moved into the companies’ internal repositories and modified their .claude/settings.json files so that the coding agent Claude Code automatically loaded the SOMBERMEME backdoor when it started. The technique did not break into Claude Code from outside; it exploited access to the repositories that the attackers had already gained. To mask its movement through the companies’ networks, the group named the GitHub personal access tokens it created after Microsoft Copilot.

GTIG does not say which setting was changed. ThreatFrontier says .claude/settings.json is the shared project settings file. It is meant to be committed to the repository, and it defines hooks: commands the agent runs automatically. A malicious change therefore spreads like any other commit. ThreatFrontier cites Claude Code’s documentation: allow rules and environment variables in the project file take effect only after the user has accepted the folder as trusted. Even so, the site advises against relying on this as the only safeguard.

ThreatFrontier says checks focus on the attack technique because GTIG published no hashes, domains, IP addresses or YARA rules for SOMBERMEME. The site advises reviewing the file’s change history across all branches, for example with git log --all -p -- .claude/settings.json. Hooks that fetch content from the internet, decode encoded content or pass commands to a shell deserve particular scrutiny. The site also recommends that a code owner always review changes to the .claude/ path, as with CI workflows. Access tokens are worth checking for names that imitate tools.

Previous coverage

Report an error in this story