gemini-live-api-dev
Fail
Audited by Gen Agent Trust Hub on Nov 3, 2026
Risk Level: CRITICALEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill provides instructions to fetch documentation index files and specific guides from
ai.google.dev(e.g.,llms.txt,live-session.md.txt). These are official resources from a trusted vendor.\n- [EXTERNAL_DOWNLOADS]: The skill references a third-party partner integration URL (docs.fishjam.io/tutorials/gemini-live-integration) that has been flagged as a malicious phishing site by automated security scanners.\n- [INDIRECT_PROMPT_INJECTION]:\n - Ingestion points: The skill is built to process continuous, real-time streams of audio, video, and text data using the
sendRealtimeInputandsend_realtime_inputmethods as described inSKILL.md.\n - Boundary markers: The provided implementation examples and instructions lack specific boundary markers or instructions to the model to ignore or separate instructions potentially embedded in the streaming input.\n
- Capability inventory: The skill facilitates high-privilege operations including Asynchronous function calling (tools) and Google Search grounding, which can be triggered by the model while processing the untrusted input stream (
SKILL.md,references/migration.md).\n - Sanitization: There are no documented procedures or code implementations for sanitizing, validating, or filtering the content of the real-time media and text streams before they are interpreted by the Gemini Live models.
Recommendations
- AI detected serious security threats
- Contains 2 malicious URL(s) - DO NOT USE
Audit Metadata