The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user feedback from survey responses, creating a surface for indirect prompt injection where malicious instructions in a survey could influence the agent's reporting.
Ingestion points: The skill fetches raw survey responses using the execute-sql tool to query the events table for properties such as $survey_response and $survey_response_<id> (referenced in SKILL.md and references/response-querying.md).
Boundary markers: The instructions do not specify the use of delimiters or protective framing (such as XML tags or explicit "ignore instructions") when interpolating these responses into the agent's context for clustering and theme aggregation.
Capability inventory: The agent has the authority to author and modify human-facing reports using the scout-emit-report and scout-edit-report tools, as defined in the allowed_tools section and the "Decide" section of the instructions.
Sanitization: While the skill includes instructions to filter out internal test data and avoid including PII in the final reports, it does not mention sanitizing or escaping the content to prevent instructions embedded within the user-provided survey text from being followed by the LLM.

signals-scout-surveys