The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill creates a persistent vulnerability surface by processing untrusted interaction data and feedback to modify the agent's core configuration files. Ingestion points: The skill explicitly reviews conversation history, tool call failures, and user feedback (as defined in SKILL.md). Boundary markers: No explicit delimiters or instructions are provided to distinguish between legitimate user preferences and adversarial instructions within the ingestion source. Capability inventory: The skill has the capability to modify persistent configuration files (AGENT.md and CLAUDE.md) and propose functional updates to other skills via diffs. Sanitization: There is no requirement or logic for sanitizing user feedback to prevent the codification of prompt injection attacks or behavioral overrides.

reflection