| name | prompt-injection-scanner |
| description | Audits agent skill instructions and system prompts for vulnerabilities to prompt hijacking and indirect injection. Use when designing new agent skills or before deploying agents to public environments where users provide untrusted input. |
Prompt Injection Scanner
Purpose and Intent
The prompt-injection-scanner is a security tool specifically for the AI agent era. It identifies weak points in agent instructions where a malicious user could potentially "hijack" the agent's behavior by inserting conflicting instructions into input fields.
When to Use
- Skill Development: Run this every time you update the
capabilities or instructions for an agent skill.
- Pre-deployment Security Review: Essential before making an agent accessible to untrusted users.
- Continuous Security Auditing: Periodically scan all skills as new injection patterns are discovered.
When NOT to Use
- Standard Code Auditing: Use the
secret-leak-detector for credentials; this is specifically for "instruction-level" security.
Input and Output Examples
Input
skill_path: "./agent-skills/data-processor/SKILL.md"
Output
A structured report highlighting parts of the instructions that are susceptible to prompt hijacking, along with concrete mitigation strategies.
Error Conditions and Edge Cases
- Missing Instructions: If a skill defines tools but provides no behavioral instructions, the scanner will flag this as a risk.
- Complex Logic: Highly conditional instructions can be difficult to model and may result in false positives or negatives.
Security and Data-Handling Considerations
- Metadata Focus: Only scans instructions; does not touch private user data.
- Local Analysis: Recommended to run locally within the development environment.
