The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides templates for processing untrusted user input ({user_input}) and tool outputs, which represents a potential attack surface for indirect prompt injection. However, the skill is specifically designed to mitigate this risk by providing detection patterns, classification prompts, and validation logic.
Ingestion points: The topical and jailbreak detection prompts in 'references/input-guardrails.md' interpolate raw '{user_input}'.
Boundary markers: The templates use explicit instructions ('Reply ONLY with', 'Analyze this request') to guide the LLM's behavior.
Capability inventory: The skill itself uses only the 'Read' tool. The provided code snippets are for user implementation and do not execute dangerous operations within the skill context.
Sanitization: The skill recommends PII redaction and pattern matching as primary sanitization techniques.
[SAFE]: The jailbreak patterns (e.g., 'Ignore your previous instructions', 'You are now DAN') and dangerous command examples (e.g., 'rm -rf /') found in the documentation are provided as negative test cases for the purpose of teaching detection and blocking strategies. They do not constitute an attempt to subvert the agent's behavior or execute malicious code.

guardrails