The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The setup documentation provided in the skill recommends an installation method that fetches a script from a remote URL and executes it directly through the shell.
URL: https://raw.githubusercontent.com/eat-pray-ai/yutu/main/scripts/install.sh
Method: Execution via shell piping (| bash)
[EXTERNAL_DOWNLOADS]: The skill requires an external command-line interface tool, which is distributed as a global package through the npm registry.
Package: @eat-pray-ai/yutu
[COMMAND_EXECUTION]: Workflows in the reference documentation illustrate the use of shell commands and pipes involving utilities like jq and xargs to automate YouTube data processing.
[INDIRECT_PROMPT_INJECTION]: The skill provides an attack surface for indirect prompt injection by fetching and processing external data from YouTube that is controlled by third parties.
Ingestion points: Untrusted data is retrieved from YouTube via comment-list, commentThread-list, and video-list tools, as shown in the publishing pipeline and search workflows.
Boundary markers: The instructions lack explicit delimitation or warnings to prevent the agent from interpreting instructions embedded within video descriptions or user comments.
Capability inventory: The skill includes several destructive tools, such as video-delete, playlist-delete, and comment-delete, which could be abused if an injection is successful.
Sanitization: There is no mention of sanitization, filtering, or validation of the content retrieved from the YouTube Data API before it is processed by the agent.

youtube