The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external text through an input field which presents a potential surface for indirect prompt injection.\n- Ingestion points: The ai_output_sample field in SKILL.md accepts untrusted text from the user which is interpolated into the model prompt.\n- Boundary markers: Absent; the prompt uses descriptive labels but does not employ strict delimiters (e.g., XML tags or triple quotes) or specific instructions to ignore commands embedded within the sample data.\n- Capability inventory: No risky capabilities are present; the skill instructions in SKILL.md are restricted to text generation and contain no network operations, file system access, or command execution logic.\n- Sanitization: Absent; the input content is interpolated directly into the prompt without validation or filtering.

ai-output-critical-audit-designer