The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill presents an attack surface for indirect prompt injection via the processing of video transcripts and user edits. 1. Ingestion points: Source transcripts are read from the project library using read_script, and user corrections are ingested via the reviewedSourceTranscript field in a UI widget. 2. Boundary markers: The untrusted data is encapsulated within <widget> and <form-textarea> tags. 3. Capability inventory: The skill utilizes submit_video_translation for processing and read_script for file access. 4. Sanitization: The instructions strictly require the application of XML attribute escaping to text placed in the widget's default attribute, which mitigates the risk of injection into the agent's context or UI.
[SAFE]: The skill's operations are limited to established vendor-specific tools and logical workflows for video localization. There is no evidence of unauthorized file access, network exfiltration, or credential exposure.

video-translation