The Agent Skills Directory

[PROMPT_INJECTION]: The skill implements an autonomous search and ingestion pipeline that is vulnerable to indirect prompt injection by design.
Ingestion points: Untrusted external data enters the agent's context through the search_tool.execute method in SKILL.md, which retrieves snippets from web search results.
Boundary markers: The skill wraps gathered results in <proactive_research> tags; while these provide a structural boundary, they do not prevent an agent from obeying adversarial instructions contained within the snippets.
Capability inventory: The agent is instructed to analyze campaign data and provide recommendations based on this injected context, allowing external content to potentially direct the agent's logic or conclusions.
Sanitization: The implementation shows no evidence of sanitization, keyword filtering, or content validation for the search results before they are formatted into the knowledge block.
[EXTERNAL_DOWNLOADS]: The skill autonomously initiates network requests to perform web searches without requiring user confirmation for each action.
Evidence: The ProactiveSearchAgent class in SKILL.md defines logic that automatically triggers and executes searches based on metric anomalies or data staleness.

proactive-intelligence