The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data retrieved from web sources to generate synthesized research reports and influence follow-up investigation strategies.
Ingestion points: Findings are retrieved from arbitrary external URLs by subagents during Phase 3 (Parallel Multi-Agent Exploration) and returned to the LeadResearcher context (SKILL.md).
Boundary markers: The skill lacks explicit instructions to treat retrieved content as data or to ignore embedded instructions (e.g., delimiters or system-level isolation are not specified).
Capability inventory: The skill utilizes the Task tool to spawn subagents and possesses file-write capabilities, saving research reports to the local file system in the .docs/ directory (SKILL.md).
Sanitization: While instructions exist to strip HTML boilerplate and advertisements, there is no mention of sanitizing or escaping the actual text content to prevent malicious instructions hidden in the data from influencing the agent's behavior or subsequent search tasks.

deep-research