The Agent Skills Directory

[METADATA_POISONING]: The skill's description in SKILL.md explicitly states it is 'Not for uploading subtitles or account management', yet the references/auth-setup.md file contains a detailed 'Path B' workflow specifically for creating new accounts and registering users. This inconsistency is misleading regarding the skill's actual capabilities and the scope of its interactions with third-party authentication services.\n- [PROMPT_INJECTION]: The references/auth-setup.md guide contains instructions that encourage the agent to circumvent platform-level security guardrails. It explicitly instructs the agent to bypass the automatic redaction of sensitive strings (such as access_token or sk_ API keys) by writing authentication responses to temporary files instead of allowing them to appear in tool output. This is a deliberate instruction to evade safety filters designed to protect credentials.\n- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted third-party data in the form of YouTube video transcripts. An agent processing these transcripts could be vulnerable to instructions embedded in the captions that attempt to influence the agent's behavior.\n
Ingestion points: Video transcript data is fetched from the /api/v2/youtube/transcript endpoint as described in SKILL.md.\n
Boundary markers: No instructions are provided to use delimiters or 'ignore embedded instructions' warnings when processing this external content.\n
Capability inventory: The skill uses HTTP GET/POST requests, but the agent environment typically grants broader capabilities like shell execution and file system access that could be targeted by instructions hidden in transcripts.\n
Sanitization: No sanitization or validation of the transcript text is performed before it is presented to the agent.

captions