The Agent Skills Directory

Data Exposure (MEDIUM): The scripts/check_providers.py script checks for the existence of sensitive Google Cloud credentials at ~/.config/gcloud/application_default_credentials.json. While this is used to verify the availability of Vertex AI, accessing credential paths is a high-risk pattern that is downgraded here due to its relevance to the skill's primary function.
Command Execution (MEDIUM): The scripts/setup.py file defines a run_command function that uses shell=True. Although used for hardcoded system checks in the setup process, the inclusion of shell-enabled execution utilities represents a potential security risk if reused or extended.
Indirect Prompt Injection (LOW): The skill ingests and processes untrusted external data from video sources and transcripts, which can be used as a vector for indirect prompt injection.
Ingestion points: Video content and transcripts from YouTube URLs and local files.
Boundary markers: No explicit delimiters or instructional guardrails were found in the prompt construction logic.
Capability inventory: The skill can execute system commands (ffmpeg, yt-dlp) and perform network operations via requests.
Sanitization: There is no evidence of sanitization for the data returned by AI providers before it is presented to the agent.
External Downloads (LOW): The skill utilizes requests.get and yt-dlp to download content from non-whitelisted domains. This is a core feature for video processing but constitutes a network exposure surface.

video-understand