The Agent Skills Directory

[DATA_EXFILTRATION]: The script scripts/fetch_and_summarize.py includes a feature to extract sensitive session cookies from the user's local web browsers (Chrome, Firefox, Edge, Safari) via the --cookies-from-browser argument. These cookies, which represent active authentication sessions, are then passed to the yt-dlp tool for use in network requests to video platforms.
[COMMAND_EXECUTION]: The script scripts/fetch_and_summarize.py uses subprocess.run to execute shell commands for installing Python packages and running the yt-dlp utility. It also utilizes __import__ for dynamic module loading, which can be an entry point for execution of untrusted code if the package names are manipulated.
[EXTERNAL_DOWNLOADS]: The skill automatically downloads and installs Python packages (yt-dlp, google-genai) from the public PyPI registry during its initial execution using the ensure_package function in scripts/fetch_and_summarize.py.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection because it processes untrusted video content from external URLs and passes it to the Gemini API for analysis and summarization.
Ingestion points: The scripts/fetch_and_summarize.py script reads URLs from user input and files, then downloads the corresponding video files from the internet.
Boundary markers: The system prompt and user-provided prompts do not include explicit instructions to the AI to ignore potentially malicious instructions embedded within the video's audio or visual data.
Capability inventory: The script has the ability to execute subprocesses, perform network operations, and write files to the local system based on the processed content.
Sanitization: While the script implements sanitize_name to clean folder names derived from video titles, the actual video content is processed raw by the AI model without validation or sanitization.

video-fetch-and-summarize