The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the creation of subagents that process potentially untrusted external data, creating a vulnerability surface. 1. Ingestion points: The subagent templates in SKILL.md (e.g., code-reviewer, debugger) are designed to ingest git diffs, codebase files, and error logs from the environment. 2. Boundary markers: The provided system prompt templates do not include specific delimiters or explicit instructions to ignore potentially malicious content embedded in the ingested data. 3. Capability inventory: The templates instruct subagents to perform actions like running 'git diff' and using the BigQuery 'bq' command-line interface. 4. Sanitization: The guidelines do not provide methods for sanitizing or escaping external content before it is interpolated into the agent's context.

create-subagent