prompt-injection-scanner

Auditing agent system prompts for prompt injection involves analyzing instruction boundaries and user input handling to identify injection vectors. This scanner reviews your SKILL.md files to surface weak points where malicious users could hijack agent behavior and provides actionable mitigations.

Indirect prompt injection occurs when untrusted user input contains conflicting instructions that hijack an agent's intended behavior. This scanner identifies these injection vectors during skill development by analyzing how your agent processes external inputs against its core system instructions.

Securing agent prompts before public deployment requires running pre-deployment security reviews on your SKILL.md instructions. The scanner analyzes instruction boundaries and user input handling to produce a risk report with recommended mitigations against instruction hijacking.

You can use this scanner for ongoing security audits of existing agent instructions and system prompts. It works with your current SKILL.md files to continuously surface potential injection vectors and provides timely guidance for secure skill design.

The best way to identify instruction hijacking risks is to analyze instruction boundaries and user input handling for injection vectors. The scanner reviews agent skill instructions to locate weak points where conflicting input could manipulate agent behavior.

You need a prompt hijacking risk assessment when designing new agent skills, conducting pre-deployment security reviews, or performing ongoing security audits. It is critical before deploying agents to public environments where users provide untrusted input.