openclaw-voice-call-realtime

Warn

Audited by Gen Agent Trust Hub on Sep 12, 2026

Risk Level: MEDIUMEXTERNAL_DOWNLOADSPERSISTENCEINDIRECT_PROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The installation process involves cloning code from a personal GitHub repository (github.com/TristanBrotherton/openclaw-voice-call-realtime) and running npm install, which presents a supply chain risk from a source not recognized as a trusted vendor.
  • [PERSISTENCE]: The skill instructions suggest configuring a Cloudflare Tunnel to run as a background service via systemd or macOS LaunchAgent, establishing a persistent network entry point that survives system restarts.
  • [INDIRECT_PROMPT_INJECTION]: The agent is exposed to untrusted voice data during phone calls, which could be used to deliver malicious instructions. 1. Ingestion points: Real-time audio stream from third-party call participants (SKILL.md). 2. Boundary markers: The skill uses talking points to define goals but lacks explicit delimiters or instructions to ignore commands embedded in the audio stream. 3. Capability inventory: The agent can autonomously execute tools to press DTMF keys, report call outcomes, and terminate the connection (SKILL.md). 4. Sanitization: There is no evidence of transcript filtering or validation to prevent the agent from acting on malicious instructions received during the call.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Sep 12, 2026, 06:50 PM
Security Audit — agent-trust-hub — openclaw-voice-call-realtime