Scan any external content before an agent INGESTS it (SKILL.md, web page, user input, tool output) for prompt-injection, data exfiltration, dangerous code execution, hidden/invisible unicode and instruction overrides: 0-100 risk score + SAFE/WARN/BLOCK verdict + structured findings with matched evidence. Is it safe to feed this content to my agent? A fast, cheap, deterministic, high-volume pre-ingest firewall. Prompt-injection / skill-audit / untrusted-content security check for agents.