The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains comprehensive documentation on how to prevent prompt injection. It describes risks like interpolating untrusted text into instructions and bypassing ingestion boundaries, but these are instructions for the agent to follow to increase security, not to bypass it.
[DATA_EXPOSURE]: No evidence of hardcoded credentials, sensitive file access, or unauthorized network operations was found. The mention of os.environ.get('PLUGIN_CACHE_PATH') is used in a defensive coding example to illustrate a path traversal risk.
[REMOTE_CODE_EXECUTION]: The skill discusses the risk of remote code execution (e.g., via exec_module or subprocess with shell=True) and provides guidelines on how to avoid these patterns. It does not contain any code that executes remote scripts.
[COMMAND_EXECUTION]: The skill provides clear instructions to avoid insecure command execution (e.g., using argv lists instead of shell=True) and to validate all inputs before they reach dangerous sinks.
[INDIRECT_PROMPT_INJECTION]: The standards/markdown-trust-surface.md file defines a robust architecture (reader/orchestrator/writer isolation) specifically designed to mitigate indirect prompt injection from untrusted data sources like web pages or GitHub issues.

plugin-security