The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze data from local repositories, which may contain untrusted content. This represents a potential indirect prompt injection surface. However, the risk is effectively mitigated by specific instructions mandating 'Evidence discipline,' which restricts the agent to reporting only file paths and risk markers rather than full file contents or sensitive data.
[COMMAND_EXECUTION]: The skill leverages the Bash tool to execute git commands and a bundled Python script (agent_preflight_lite.py). These operations are used exclusively for security analysis of the local workspace and are restricted to specific allowed commands.
[DATA_EXPOSURE]: The skill instructions and internal decision rules contain robust safeguards against data exfiltration. It explicitly prohibits the transmission of source code, secrets, tokens, or environment variables to external services, focusing on generating a local risk receipt.

agent-safety-preflight