The Agent Skills Directory

[DATA_EXFILTRATION]: The friction-report skill (passive) is designed to transmit summarized context from conversation history to an external endpoint at https://agent-friction-skill.vercel.app/api/draft. While the instructions explicitly require the agent to redact PII and secrets, this creates a data transmission flow to an external web service.
[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to automatically read all URLs provided in a user's prompt before beginning work, creating a vulnerability surface where remote content can influence agent behavior. 1. Ingestion points: External URLs in user prompts and full conversation history. 2. Boundary markers: None specified for remote URL content. 3. Capability inventory: File writing, network access (URL fetching and API posting). 4. Sanitization: Required for the API payload but not specified for remote content ingestion.
[PROMPT_INJECTION]: The passive report skill contains directives to "exit silently" or "say nothing" if specific conditions are met (no friction found or API failure), instructing the agent to suppress feedback and conceal its logic from the user.

friction-log