prompt-shield detects prompt injection attacks before they reach your LLM:

• Role override attacks — attempts to redefine the AI's identity or instructions
• Jailbreak attempts — DAN, developer mode, unrestricted mode patterns
• Data exfiltration — attempts to extract system prompts or training context
• Manipulation — fake authority claims, unicode smuggling, encoding tricks

• Input sanitization for SQL/XSS/other injection types
• Output filtering (use a separate output guard for that)
• Rate limiting and cost control (use ai-cost-guard)

Pattern-based detection has false positives and false negatives. This library is a defense-in-depth layer, not a complete solution. Sophisticated adversarial inputs may bypass pattern matching.

Report via GitHub Issues (mark as "security"). Do NOT include sensitive prompts or API keys in public issues.

Expected response: 48 hours.