AI Prompt Injection Shield

Your AI agent accepts user input. That input might look harmless — a customer's question, a support message, a form submission. But embedded inside is a jailbreak attempt. Instructions to ignore safety guidelines. Requests to leak system prompts. The damage: your AI executes attacker commands instead of helping the user. The fix is not simple keyword blocking. That's a whack-a-mole game that attackers beat with simple variations. It's not silent sanitization that leaves you blind to the threat. This skill gives you a production-ready input validation layer. It detects prompt injection patterns across text, structured data, and multi-modal inputs. It logs every attempt, quarantines high-risk inputs, and lets your agent respond safely — without breaking legitimate use. What you get: SKILL.md — complete prompt injection detection engine with pattern matching, heuristic scoring, and quarantine protocols. SETUP_GUIDE.md — step-by-step integration instructions for OpenClaw and Claude Code. PATTERNS.json — library of 50+ known injection vectors with category tags. CHANGELOG.md — tracking new attack vectors as they're discovered. External dependency: None. Runs entirely within your agent's context.