The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external codebases and Pull Requests, creating an attack surface for indirect prompt injection. Ingestion points include git clone and gh pr checkout to bring external code into the agent's environment. The skill lacks boundary markers or sanitization to isolate external code from system instructions.
[DYNAMIC_EXECUTION]: The skill explicitly advises using flags like --yolo for Codex, which is described as bypassing security sandboxes and human review. It also uses tmux send-keys to programmatically interact with sessions.
[EXTERNAL_DOWNLOADS]: Recommends installing a global npm package @mariozechner/pi-coding-agent from a source not verified as a trusted vendor.
[COMMAND_EXECUTION]: The skill provides a framework for background execution of arbitrary agent-generated commands through various CLI tools.
[DATA_EXFILTRATION]: The skill instructions acknowledge that agents can access sensitive local files like soul.md or the ~/clawd/ directory, and provides patterns to post results to external services like GitHub.

coding-agent