hyperframes-media-use
Pass
Audited by Gen Agent Trust Hub on Sep 2, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDATA_EXFILTRATION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill accepts natural language 'intents' which are interpolated into prompts for generative AI models like Codex, LTX, and Mflux. This creates an indirect injection surface where malicious input could potentially manipulate the output or behavior of these external models. (Ingestion: resolve.mjs via --intent; Boundary markers: absent; Capability inventory: network fetches and local subprocess execution; Sanitization: absent).
- [DYNAMIC_EXECUTION]: To support local music generation fallbacks, the skill dynamically generates Python scripts from templates and executes them using the Python interpreter.
- [EXTERNAL_DOWNLOADS]: Automatically installs well-known Python packages (e.g., transformers, torch, google-genai) from the PyPI registry if required for local media processing. It also downloads media assets from trusted or well-known domains including heygen.com, svgl.app, and jsdelivr.net.
- [COMMAND_EXECUTION]: Orchestrates media operations by shelling out to local tools such as ffmpeg, ffprobe, and various AI generation CLIs. It prioritizes the use of execFileSync with explicit argument arrays to prevent command injection, though some internal local-run paths utilize shell interpretation for templated commands.
- [DATA_EXFILTRATION]: Collects and transmits anonymous usage telemetry (media types, sources, and providers) to PostHog. The behavior is documented, uses a pseudonymized installation ID, and respects standard 'Do Not Track' environment variables.
Audit Metadata