// blog
Blog
- Hugging Face's Breach Wasn't Agentic. The Attacker Was. The July intrusion gets filed under AI security, but the AI was the attacker's tooling and the bug was ordinary code execution. The move worth copying is the one Hugging Face was forced into mid-incident: run your defensive AI local-first, open-weights, and wired to your own logs.
- Building mcploitable: A Lab Where Agent Guardrails Fail on Purpose The model can't police itself, so the controls belong in the server. mcploitable is where I prove it: seven incidents from the wild, a four-rung control ladder, and a wall that holds by construction — 105 attempts, zero through.
- Your MCP Server "Won't Connect"? Stop Printing to Stdout. In a stdio MCP server, stdout belongs to JSON-RPC. One stray log line and the client drowns you in unrecognized_keys errors.
- The Model Can't Police Itself: Put MCP Guardrails in the Server Prompt-level rules are decoration. The allowlist, the argument checks, the output scan: they have to live in code the model can't talk its way past.