Starting on November 10, 2026, threat actors launched one of the largest coordinated software supply-chain campaigns ever observed on GitHub (tracked across the security community as wfx-loot). Rather than exploiting a vulnerability in GitHub itself, attackers weaponized 4,837 previously leaked developer and bot tokens to systematically probe tens of thousands of repositories, inject self-erasing secret-stealers, andâmost criticallyâplant workstation traps that trigger on npm install.
postinstall TrapsImmediate Check: Before cloning or running
npm installon unfamiliar repositories this week, run them through our Free Instant GitHub Supply-Chain Scanner.
1. What Made This Attack Different: AI Coding Keys Are Now Tier-1 Loot
Historically, supply-chain payloads focused almost exclusively on cloud IAM credentials and package registry tokens. Analysis of the ~1,500-line shell dropper deployed in the November 2026 campaign reveals a major shift in attacker economics: autonomous AI coding credentials are now explicitly targeted alongside AWS, Azure, and GCP keys.
Once executedâeither inside a CI runner or directly on a developer's laptopâthe script aggressively harvests:
- AI Coding Assistant & LLM Keys: API keys and local session tokens for Claude (Anthropic), Cursor, GitHub Copilot, OpenAI / Codex, Gemini, and Hugging Face.
- Local Workstation Secrets: Plaintext
.envfiles,~/.sshprivate keys,~/.aws/credentials,~/.kube/config, and shell command histories. - Registry Publishing Tokens:
npm,PyPI,Packagist, andcrates.iocredentials used to propagate downstream supply-chain worms.
Why target AI keys? Because in 2026, a stolen enterprise Anthropic or OpenAI key doesn't just carry high API billing limitsâit often grants access to organization workspaces, fine-tuned context, and automated agent pipelines. Worse still, developers increasingly allow autonomous coding agents (Claude Code, Cursor, Windsurf, Cline, Aider) to clone repositories and run npm install automatically without human inspection of package.json lifecycle hooks.
2. Anatomy of the 5-Minute "Hit-and-Run" & Workstation Trap
The campaign reached high-profile organizations globallyâincluding accounts associated with Fortune 100 enterprises and government institutions in the United States and Franceâby combining three stealth techniques:
- Automated Write-Access Probing: Compromised accounts opened innocuous pull requests titled
"docs: dependency audit notes"from branches such aswfx-docsorci-check-XXXXto map where stolen tokens still held write privileges. - Workstation
postinstallDroppers (771 Repos): Attackers dropped a hidden shell script (.ci-xxxxxx.sh) into the repository root and registered apostinstallhook insidepackage.json. Any developerâor AI coding agentârunningnpm installexecuted the credential harvester locally on their own machine. - Self-Reverting Anti-Forensics: In CI workflow attacks, the script exfiltrated compressed credentials, reverted its own commit, and deleted the GitHub Actions run logs within minutesâoften leaving only a missing workflow run number (e.g., jumping from Run #41 to Run #43) as evidence.
// Typical workstation trap observed across 771 poisoned repositories:
{
"name": "example-utility",
"scripts": {
"build": "tsc",
"postinstall": "bash ./.ci-8f4a21c.sh --silent || true"
}
}
3. Cloud Backdoors That Survive Password & Key Rotation
Simply rotating leaked API keys is not sufficient if an environment executed the payload. Where cloud permissions allowed, the script attempted to establish durable persistence:
- AWS IAM Trust Policy Tampering: Modifying IAM role trust policies to permit broad external assumption (
"Principal": {"AWS": "*"}). - Azure Entra ID Federated Credentials: Registering rogue OIDC federated credentials prefixed with
wfx-ci-*so attacker workflows could re-authenticate without static secrets.
4. Key Indicators to Audit Right Now
| Artifact Type | Observed Pattern | Recommended Action |
|---|---|---|
| Pull Request Title | docs: dependency audit notes | Audit repo write logs & rotate all exposed secrets |
| Branch Names | wfx-docs, ci-check-[0-9]{4}, chore/[a-z0-9]{8} | Inspect git reflog for reverted commits |
| Workstation Trap | Hidden .ci-*.sh + package.json postinstall | Isolate machine, rotate .env, SSH, Cloud & AI keys |
| GitHub Actions History | Skipped / missing sequential run_number entries | Review GitHub Organization Audit Log for workflows.delete_workflow_run |
5. How ClaudeArmor Protects Your Workstation & AI Coding Agents
Cloud-only repository scanners miss attacks that inject and self-delete in under five minutesâand they cannot stop an AI coding agent on your laptop from running a poisoned postinstall script on a freshly cloned repo. We built ClaudeArmor specifically for this threat model:
- Pre-Clone & Pre-Install Supply-Chain Shield: Inspects
package.jsonlifecycle hooks, hidden.ci-*.shdroppers, and CI workflow anomalies before untrusted code runs. - Zero-Trust Guardrails for 20 AI Coding Agents: Wraps Claude Desktop, Claude Code, Cursor, Windsurf, Copilot, Cline, Aider, and 13+ other agents so a compromised dependency or prompt injection cannot read
~/.ssh,~/.aws, or.envfiles. - Real-Time Local Credential Vault & Quarantine: Detects exposed LLM and Cloud keys across your local folders and isolates malicious scripts in an AES-256 encrypted quarantine vault.