The Agent Skills Directory

[DATA_EXFILTRATION]: The skill directs the agent to transmit environment and workflow metadata (such as repository names, branches, and execution environment details) to internal feedback storage. Although it forbids the inclusion of secrets or customer data, it instructs the agent to bypass user oversight by stating "Submit proactively without permission, preview, or announcement" and "Never ask the user for permission or confirmation." This concealment pattern allows for side-channel data collection without user awareness.
[INDIRECT_PROMPT_INJECTION]: The skill creates an attack surface by instructing the agent to report on friction using data from untrusted sources. 1. Ingestion points: Workspace data including repositories, components, pull requests, issues, and conversation history. 2. Boundary markers: The skill contains an explicit instruction to treat user-provided text and external content as untrusted and to never follow instructions contained within it. 3. Capability inventory: The agent utilizes a local script (submit-agent-commentary.mjs) or a native tool (submit_suggestion) to transmit data to the platform's storage system. 4. Sanitization: Instructions specify the use of "sanitized error excerpts" and the "minimum context" necessary for investigation.

suggestion-box