voice-of-customer-synthesizer

Pass

Audited by Gen Agent Trust Hub on Jul 20, 2026

Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: The skill processes untrusted external data from public review sites and social media platforms, creating a surface for indirect prompt injection.
  • Ingestion points: External data is collected in Phase 1B (Review scraping) and Phase 1C (Social listening) from platforms like G2, Reddit, and Twitter.
  • Boundary markers: The instructions lack specific delimiters or "ignore embedded instructions" warnings to prevent the agent from obeying malicious commands hidden within customer feedback.
  • Capability inventory: The skill includes the capability to write the synthesized report to the local file system (Phase 5).
  • Sanitization: No sanitization, validation, or escaping logic is described for the incoming text data before it is processed by the LLM for theme clustering and recommendations.
  • [COMMAND_EXECUTION]: The documentation includes an example of a cron job for automation that executes a shell command (python3 run_skill.py). The referenced script run_skill.py is not included in the provided skill files, making its exact behavior unverifiable.
Audit Metadata
Risk Level
SAFE
Analyzed
Jul 20, 2026, 06:25 PM
Security Audit — agent-trust-hub — voice-of-customer-synthesizer