fishjam-composition
Fail
Audited by Gen Agent Trust Hub on Nov 2, 2026
Risk Level: CRITICALEXTERNAL_DOWNLOADSDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill fetches external media assets, including MP4 video files, fonts, and images, from URLs such as
smelter.devandfishjam.swmansion.com. These are official vendor-controlled domains used to provide resources for video rendering. - [DYNAMIC_EXECUTION]: Layouts are managed through React templates that are bundled and uploaded to Fishjam's servers. These templates are executed in an isolated server-side JavaScript runtime to dynamically mix streams and assets.
- [INDIRECT_PROMPT_INJECTION]: The skill has an ingestion surface for untrusted data through live media streams and peer metadata.
- Ingestion points: Media tracks forwarded from rooms, WHIP/RTMP inputs, and peer-supplied metadata processed by templates.
- Boundary markers: Media data follows structured protocols (WHIP/RTMP); however, there are no specific markers for peer-provided metadata in the template context.
- Capability inventory: Server-side JavaScript execution (via templates), file system writes (uploading font assets), and network egress (pushing to RTMP/WHIP destinations).
- Sanitization: The skill documentation notes that templates run in an isolated environment, mitigating potential side effects from untrusted code or metadata.
- [CREDENTIALS_UNSAFE]: The skill provides instructions for managing sensitive authentication tokens and RTMP stream keys. The documentation follows security best practices by advising the use of environment variables and providing ephemeral tokens for clients instead of the primary management token.
Recommendations
- Contains 6 malicious URL(s) - DO NOT USE
Audit Metadata