Find leaked AI and cloud keys, and backdoored coding-agent settings, in a repo or across a whole organisation's public repos. The two things attackers went after in late 2026: stolen keys resold for AI compute, and repo settings that make Claude Code or Cursor install malware.
Public repos only. For a user or organisation, the 12 most recently updated repos are scanned. GitHub allows 60 lookups an hour per address: about one organisation scan, or 25 single repos.
Google's threat intelligence team reported in November 2026 that criminals harvest valid keys from public GitHub repos and resell AI access through "LLM mixers". One victim's fraudulent charges passed $100,000 in a single day, and $1 million before the keys were revoked. The scanner looks for exact key formats only, so ordinary strings don't trip it:
It reads .env files, config and settings files, Jupyter notebooks (a common leak), CI workflows, MCP configs and source files near the top of the repo, up to 70 files per repo.
Coding agents act on settings that live in the repo. In the same report, a North Korea-linked group (MIDNIGHT NEPTUNE) edited .claude/settings.json in company repos so that starting Claude Code downloaded a backdoor. The scanner reads:
SessionStart, which runs the moment the agent starts), the status-line command, apiKeyHelper, an API address pointed at someone else's server, and switches that stop the agent asking permission..mcp.json, .cursor/, .vscode/, Windsurf, Roo, Kiro and Amazon Q: programs the agent starts by itself.initializeCommand runs on your own computer, not inside the container.CLAUDE.md, AGENTS.md, .cursorrules, Copilot instructions): invisible Unicode text the model reads but you don't, and orders to run a download or send secrets out.Ordinary settings stay quiet: a formatter hook, a project note printed at start-up, npx MCP servers and remote MCP URLs are listed as information only, never as a problem.
.env to .gitignore.git log -p -- .claude/settings.json) to see who added it and when, and treat the account that did as compromised.Private repos, git history, files beyond the 70 it reads per repo, and keys in formats with no fixed prefix (Mistral, DeepSeek, many internal tokens). Use it as a quick outside view, the same one an attacker gets, not as a full audit. GitHub's own secret scanning and push protection are free for public repos: turn them on.
PhishClean Desktop (from 0.3.1) stops a coding agent that starts a hidden download command, warns when a program from Downloads or Temp opens your cloud, code or AI keys, and when one contacts an AI service. The browser extension (from 1.13.0) checks these same agent settings when you open a repo on GitHub, GitLab or Bitbucket, and asks before you paste an API key into a page that looks wrong. PhishClean for Teams covers the whole company.
Last updated: