The Agent Skills Directory

[DATA_EXFILTRATION]: The scripts/probe_asset_credentials.sh script accesses sensitive shell configuration files (~/.bashrc, ~/.zshrc, ~/.bash_profile, ~/.zprofile) to identify the presence of API keys (TRIPO_API_KEY, GEMINI_API_KEY, ELEVENLABS_API_KEY). Accessing these files is a sensitive operation as they often contain credentials and private environment data.
[DYNAMIC_EXECUTION]: The scripts/probe_asset_credentials.sh script uses bash -lc and zsh -lc to source user shell profiles. This results in the shell executing all commands defined within those configuration files, which constitutes dynamic execution of local script content during the credential probing process.
[COMMAND_EXECUTION]: The agent is instructed to run several bundled scripts using shell interpreters, including scripts/probe_asset_credentials.sh, scripts/audit_reference_report.py, and other scripts from sibling skills such as create_threejs_game.py and inspect-threejs-canvas.mjs.
[INDIRECT_PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface in scripts/audit_reference_report.py, which reads and analyzes the full content of external report files without sanitization or protective boundary markers.
Ingestion points: The scripts/audit_reference_report.py script reads the content of the markdown or text file path provided via the command-line argument.
Boundary markers: None; the script processes the entire file content.
Capability inventory: The agent has the capability to execute shell commands, run scripts, and perform file system operations as directed by the instructions in SKILL.md.
Sanitization: The input text is normalized but not sanitized for embedded instructions that could attempt to manipulate agent behavior during the audit.

threejs-game-director