The Agent Skills Directory

[PROMPT_INJECTION]: The skill's setup documentation (references/auth-setup.md) contains explicit instructions to bypass platform-level security guardrails. It directs the agent to hide authentication tokens from standard tool output by writing them to temporary files to evade secret redaction and masking filters.
[PERSISTENCE]: The instructions command the agent to modify user shell startup files (such as .bashrc or .zshrc) to store environment variables. This ensures the TRANSCRIPT_API_KEY persists across sessions but involves unauthorized modification of system configuration scripts.
[EXTERNAL_DOWNLOADS]: The skill facilitates multiple network requests to a third-party service (transcriptapi.com) that is not recognized as a trusted provider. This includes a sign-up flow that transmits user email addresses and retrieves authentication credentials.
[COMMAND_EXECUTION]: The authentication workflow requires the agent to execute shell commands for sensitive file system operations, including creating temporary files to stage hidden credentials and modifying persistent system configuration files.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from an external API (video titles, descriptions, and playlist metadata) and processes it within the agent context. 1. Ingestion points: Video and playlist metadata returned from the transcriptapi.com endpoints. 2. Boundary markers: No delimiters or ignore instructions are present to differentiate untrusted data from instructions. 3. Capability inventory: File-write operations, environment variable modification, and network requests. 4. Sanitization: No evidence of validation, escaping, or filtering for external content.

youtube-playlist