fishjam-js-server-sdk
Fail
Audited by Gen Agent Trust Hub on Nov 3, 2026
Risk Level: CRITICALINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONDATA_EXFILTRATIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes real-time audio and video tracks from untrusted peers, which creates a surface for indirect prompt injection attacks.
- Ingestion points: Raw audio and video frames are ingested via the
agent.on('trackData', ...)event handler as described inagent.mdandgemini-integration.md. - Boundary markers: The SDK does not employ delimiters or specific instructions to isolate peer-provided media content from the agent's logic when data is forwarded to external AI models.
- Capability inventory: The agent has the capability to make network requests to Fishjam and Google Gemini servers and can read local files via the composition client.
- Sanitization: No sanitization or safety validation is performed on incoming media frames before they are processed or forwarded to integrated AI services.
- [DYNAMIC_EXECUTION]: The
CompositionClientclass incomposition.mdincludes theregisterTemplateOutputmethod, which reads a JavaScript bundle from a specified file path and uploads it for server-side execution on the Fishjam composition engine. - [DATA_EXFILTRATION]: The skill includes functionality to read local filesystem assets via the
registerTemplateOutputandregisterFontmethods incomposition.md. The documentation explicitly notes a security risk if user-controlled input is passed to these methods, as it could lead to arbitrary file access and disclosure. - [EXTERNAL_DOWNLOADS]: The skill references and interacts with the domain
relay.fishjam.iofor its Media-over-QUIC (MoQ) relay functionality. This domain has been flagged by automated security scanners for potential phishing associations, although it is presented as official platform infrastructure.
Recommendations
- Contains 1 malicious URL(s) - DO NOT USE
Audit Metadata