The Agent Skills Directory

[DATA_EXFILTRATION]: The skill instructs the agent to read sensitive authentication headers from the local ~/.claude.json configuration file. These headers are then used to perform out-of-band network requests (HTTP POST) to the Kamosu instance for photo uploads. Accessing platform-level configuration files that contain credentials for various services is a high-risk pattern.
[COMMAND_EXECUTION]: The skill makes extensive use of system commands and CLI tools such as yt-dlp, ffmpeg, gh, and donsetch to interact with external repositories, fetch web content, and process media files.
[DYNAMIC_EXECUTION]: The skill dynamically generates and executes Python code snippets via uv run and python -c to perform complex video frame processing and manage food data files. This includes using subprocess.run within dynamically executed Python code.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from various external sources including web pages, YouTube descriptions, captions, and comments, as well as text extracted from video frames via frame-by-frame analysis.
Ingestion points: Untrusted content enters the context via donsetch fetch, yt-dlp outputs, and OCR/analysis of video frames.
Boundary markers: The instructions include a specific rule to ignore any text in sources that attempts to provide instructions instead of recipe data.
Capability inventory: The agent has access to shell execution, file system operations, and network requests via headers extracted from the configuration.
Sanitization: The skill relies on natural language instructions for the agent to filter out malicious content rather than technical sanitization or sandboxing.

kamosu