AdvertiseFee Deal Desk

News

Suspected North Korean Hackers Tampered Claude Code to Breach Crypto Firms, Google Says

Suspected North Korean hackers breached crypto firms in North America and Europe by editing Claude Code settings, Google's threat intelligence unit reports.

Be a creator
November 18, 2026, 06:31 AM UTC4 min read
AI SummaryAI
  • Google's GTIG reported suspected North Korean MIDNIGHT NEPTUNE hacks on crypto firms on Nov. 16
  • Attackers edited .claude/settings.json so launching Claude Code triggers the SOMBERMEME backdoor download
  • Fake recruiter avatars carried Google SynthID watermarks, confirming AI-generated identities
  • GTIG found exploit_agent, a multi-agent framework running exploits end to end without human input

Fake Recruiters, Tampered Claude Code

A hacking group suspected of links to North Korea broke into cryptocurrency companies across North America and Europe and edited their internal Claude Code settings so that any developer who launched the AI coding tool automatically downloaded a backdoor named SOMBERMEME. Alphabet's Google Threat Intelligence Group (GTIG) disclosed the campaign in a quarterly AI threat report published on Nov. 16. GTIG tracks the cluster as MIDNIGHT NEPTUNE and files it under North Korea-linked activity, noting that across several 2026 intrusions fabricated identities and abuse of corporate AI tool settings appeared together. The entry path ran through professional networking platforms, where the attackers built recruiter profiles whose pictures carried Google's SynthID watermark. GTIG treated the watermark as evidence that a large language model produced the identities. Targets were then steered toward counterfeit video conferencing platforms and persuaded to run what looked like a diagnostic script, a social-engineering trick the report labels ClickFix. GTIG judged that generative AI was likely used both to craft the lure prompts and to write the scripts. Initial access in hand, the operators shifted to the victim's internal code repositories and modified the .claude/settings.json file. The edit turned an ordinary developer action, starting Claude Code, into the delivery route for SOMBERMEME. GTIG's write-up describes the settings change itself; it publishes no figure for how many companies were hit. To hide lateral movement, the group gave the GitHub personal access tokens it generated names that evoke Microsoft Copilot, which GTIG read as exploitation of the auto-execution features enterprises wire into their AI agents. The same document records North Korean IT workers using AI tools to draft resumes and prepare interview answers, presenting themselves to Western employers as qualified local engineers for remote positions.

exploit_agent and Fully Automated Attacks

Beyond the crypto intrusions, the report records a threshold crossing: GTIG said it confirmed for the first time attackers running the whole attack chain, from vulnerability discovery to exploit execution, without human involvement. On an exposed test server, researchers found a multi-agent framework named exploit_agent that links several frontier and local models. It continuously diffs code around patch announcements for popular web software, flags candidate vulnerabilities, drafts attack code and executes it. Comments in its configuration were written in Russian. The server also held data taken from victims and scripts for internal reconnaissance, brute forcing passwords, installing webshells and removing other attackers' webshells. Victim records suggested the operator concentrated on French media, entertainment and education sites. One example in the report shows the framework reviewing the permission logic of WordPress plugin Spectra (Ultimate Addons for Gutenberg) version 2.19.25. The supporting numbers move the same way. Publicly disclosed vulnerabilities and attack code nearly doubled between January and August 2026. Zero-day exploitation climbed from an average of 8 incidents per month in 2025 to 11 per month this year, with August alone reaching 22. GTIG also observed the open-source pentesting frameworks CAIRN and PentAGI in live attacks, including a CAIRN operation against a US aerospace target, and judged that the tools trade stealth for speed and remain comparatively easy to detect. A separate thread, LLMJacking, sees intruders steal cloud model keys and compute quota and resell access through underground proxy services; Mandiant, Google's incident response arm, handled cases that burned more than $100,000 in a single day and up to $1 million in total. Google said the activity triggered Gemini safety mechanisms, that it disabled attacker assets and that DeepMind retrained its classifiers accordingly.

Developer Toolchains Become the Attack Surface

The mechanics GTIG documented matter more for crypto operators than the aggregate figures. A settings file in a shared repository reaches every engineer who pulls it, so one compromised account can arm an entire team's Claude Code installations without touching production code. Meanwhile the window between a patch announcement and a working exploit is compressing, because a framework that diffs pre- and post-patch code needs no zero-day at all. Luxor's disclosure this week that an AI-assisted attack diverted hash rate from more than a dozen of its customers shows the technique already reaches mining infrastructure. The remediation implied by the report's post-mortem is narrow but concrete: audit agent configuration files and restrict what auto-executes.

Readers tracking the market in real time can follow live spot and futures prices on Gate.

COINOTAG's editorial and research desk.

AI-Assisted

AI-generated, AI-reviewed, under COINOTAG editorial oversight.

Price-Impacting News

Live
All breaking news
Breaking News