What it is:
- Go framework exposing decoy tools over MCP that look legitimate (names/params/descriptions), return safe dummy output, and emit telemetry when invoked.
- Runs alongside your real tools; ship events to stdout/webhook or your pipeline (Prometheus/Grafana, ELK).
Why it helps:
Agent logs show what happened; canaries mark what must not happen. A single tripwire is an immediate, low-noise indicator of compromise.
Real-world relevance (Nx attack):
Recent reporting on the Nx npm supply-chain incident (“s1ngularity”) shows malicious versions exfiltrated SSH keys, tokens, and other secrets—and notably abused AI developer tools like Claude/Gemini in the workflow, one of the first documented cases of AI assistants being weaponized in a software supply-chain attack. If your IDE agent (Claude Code or Gemini Code/CLI) had a canary tool registered—e.g., a fake “export secrets” or “repo exfil” action—any unauthorized tool call from the agent side would have triggered a deterministic alert during that incident.
Links:
GitHub: https://github.com/mariocandela/beelzebub
Blog: https://beelzebub-honeypot.com/blog/securing-ai-agents-with-...
Feedback wanted! :)