The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The script get_audio.py is described as automatically downloading and updating the yt-dlp utility when needed. While yt-dlp is a well-known tool, the automated download and execution of remote binaries at runtime is a high-privilege operation.\n- [COMMAND_EXECUTION]: The skill makes extensive use of system shell commands to orchestrate Node.js scripts, Python scripts, ffmpeg, and Chromium. This requires the agent to have broad execution permissions in the local environment.\n- [CREDENTIALS_UNSAFE]: The skill is designed to read OPENAI_API_KEY or ELEVENLABS_API_KEY from ~/.config/gpt-js-animations/keys.env. Although the instructions emphasize keeping these private, the agent is granted intentional access to sensitive configuration files containing API keys.\n- [DYNAMIC_EXECUTION]: The core workflow involves the agent writing JavaScript and HTML code (film/index.html) based on a 'treatment' and then executing that code via a headless browser to generate video frames. This 'generate-then-execute' pattern is a primary vector for potential exploitation if the generation logic is compromised.\n- [INDIRECT_PROMPT_INJECTION]: The skill has a significant attack surface for indirect prompt injection as it processes external content from URLs (via yt-dlp) and transcribes audio (via openai-whisper).\n
Ingestion points: External audio/video URLs provided by the user; local audio files processed for transcription.\n
Boundary markers: No specific boundary markers or sanitization logic is described for the content being ingested.\n
Capability inventory: Extensive capabilities including file writes, network downloads, and shell command execution.\n
Sanitization: No explicit sanitization or validation of the input data is mentioned in the skill instructions.\n- [EXTERNAL_DOWNLOADS]: The skill fetches the yt-dlp executable and openai-whisper model weights from remote sources. While these are well-known services and tools, the skill initiates these downloads automatically during operation.\n- [METADATA_POISONING]: The documentation references non-existent models such as 'GPT-6 Astra' and 'GPT-6.1 Sol'. This misleading metadata could cause a user or agent to misinterpret the required environment or the skill's actual capabilities.

gpt-js-animations