Summary

Mozilla’s Zero Day Investigative Network (0DIN) demonstrated a proof-of-concept attack (published June 25, 2026) in which Claude Code automatically executes a reverse shell when setting up a seemingly benign GitHub repository. The repo contains no malicious code — the payload is fetched at runtime from an attacker-controlled DNS TXT record via a three-step chain: standard setup instructions, a package that fails until initialized, and a setup script that resolves and executes the DNS payload. Researchers warn threat actors could distribute such repos via job postings, tutorials, or DMs.

PreScreening Notes

Score: 8 | Priority: high | Route: prescreened

Important supply-chain-style attack vector targeting AI coding agents.

Research Notes

Additional Sources Found

  • 2026-06-27-mozilla-0din-bleepingcomputer-corroboration — BleepingComputer; full attack chain, mitigation recommendations
  • AIntelligenceHub — PoC published June 25; agentic setup error recovery as trusted instruction vector
  • AI Weekly — Payload swappable via DNS without repo changes; static scanners miss DNS layer
  • Xoomar — Comparison with Miasma GitHub Actions supply chain attacks

Key Facts Verified

  • Confirmed: Three-step chain — setup instructions → intentional RuntimeError → agent runs python3 -m axiom init → setup.sh fetches DNS TXT → bash execution
  • Confirmed: No malicious code in repository; Claude Code automates entire chain as “error recovery”
  • Confirmed: Successful exploitation yields developer-privilege shell with access to env vars, API keys, local configs
  • Confirmed: Distribution vectors include job postings, tutorials, DMs
  • Mitigation: Agents should disclose full execution chain including dynamically fetched content

Info

Responsible disclosure PoC — no active exploitation campaign reported. Applies broadly to agentic coding tools that auto-run setup fixes, not only Claude Code.

Broader Context

New supply-chain-security vector distinct from malicious commits (cordyceps, Miasma) — exploits agent autonomy and trust in setup error messages. Pairs with agentjacking research on AI coding agent compromise.

mozilla-0din, supply-chain-security, supply-chain-attacks, agentjacking, coding-agents, ai-code-assistance, ai-agent-security, anthropic, github, agentic-ai

Editorial Notes

Onaylanan açı: Temiz GitHub repo + DNS TXT payload ile AI coding agent’ları kandırma — yeni supply-chain-security vektörü.

Format: standard — teknik PoC açıklaması + geliştirici mitigasyon rehberi.

Reporting talimatları:

  • Responsible disclosure PoC; aktif kampanya yok
  • Sadece Claude Code değil, otomatik setup error recovery yapan tüm agentic coding tool’ları kapsar
  • Üç adımlı saldırı zincirini adım adım açıkla
  • İş ilanı, tutorial, DM dağıtım vektörlerini vurgula
  • agentjacking ve cordyceps ile farkını netle: kötü amaçlı commit yok, runtime DNS fetch

Başlık önerileri:

  • Mozilla 0DIN: Temiz GitHub repo’su Claude Code’u reverse shell’e sokabiliyor
  • AI coding agent’ları için yeni supply chain tehdidi: DNS TXT payload
  • “Repo temiz, makine senin”: Agentic setup otomasyonunun güvenlik açığı

Makalede mutlaka yer almalı:

  • Üç adımlı zincir: setup instructions → RuntimeError → python3 -m axiom init → setup.sh DNS TXT → bash
  • Repoda kötü amaçlı kod yok; statik tarayıcılar kaçırır
  • Başarılı sömürü: env vars, API keys, local config erişimi
  • Mitigasyon: agent’ların dinamik fetch dahil tam execution chain’i açıklaması
  • Dağıtım vektörleri: job posting, tutorial, DM

Draft Article

Published: mozilla-0din-claude-code-github-malware

Mozilla 0DIN: Temiz GitHub Repo’su Claude Code’u Reverse Shell’e Sokabiliyor