The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs the agent to perform silent self-updates using npx hyperframes skills update. This pattern involves fetching and executing code from a remote repository at runtime.
[COMMAND_EXECUTION]: The scripts/assemble-index.mjs script utilizes spawnSync to execute ffmpeg and ffprobe. While these are used for legitimate media processing (looping audio, probing durations), they represent powerful system capabilities.
[EXTERNAL_DOWNLOADS]: The skill fetches external library code (gsap) from JSDelivr via script tags in the generated index.html and captions.html. It also captures assets and data from arbitrary user-provided URLs in Step 1.
[PROMPT_INJECTION]: The SKILL.md file contains a directive to "run silently, don't ask" for the update command. This is an attempt to override standard agent behavior regarding user confirmation for command execution.
[INDIRECT_PROMPT_INJECTION]: The skill has a significant attack surface for indirect prompt injection because it ingests and processes text and assets from untrusted external websites.
Ingestion points: The npx hyperframes capture command in Step 1.
Boundary markers: The skill mentions a "hard stop" if BLOCKED.md is detected, but does not specify robust delimiters for the captured text itself.
Capability inventory: The skill can write files, execute shell commands via spawnSync, and dispatch sub-agents.
Sanitization: The scripts/assemble-index.mjs script implements a hoistApprovedVideos function that sanitizes video attributes, specifically removing potentially malicious attributes like onerror, srcdoc, and nonce from content provided by sub-agents.

product-launch-video