The Agent Skills Directory

[PROMPT_INJECTION]: The skill presents an attack surface for indirect prompt injection because its core function is to process data from unverified external sources. \n- Ingestion points: Untrusted content enters the agent context through the various research sources used by subagents to compile findings (SKILL.md). \n- Boundary markers: The instructions do not specify delimiters or clear instructions for subagents to ignore or bypass embedded directives found in retrieved content. \n- Capability inventory: The skill possesses the ability to spawn subagents, create directories, and write multiple Markdown files to the local file system (SKILL.md). \n- Sanitization: No sanitization or validation mechanisms are described to filter potentially malicious instructions from external sources before processing. \n- [NO_CODE]: The skill consists exclusively of instructions in Markdown format and does not include any accompanying scripts or binary files, which minimizes the risk of direct malicious code execution or persistence.

research