openai-whisper-hardened
Pass
Audited by Gen Agent Trust Hub on Apr 21, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADS
Full Analysis
- [SAFE]: The skill implements strong safety boundaries by instructing the agent to never exfiltrate transcribed text to network endpoints, ensuring local-only processing.
- [EXTERNAL_DOWNLOADS]: The skill installs the openai-whisper package via Homebrew, which is a well-known and trusted utility from a trusted organization (OpenAI).
- [COMMAND_EXECUTION]: The skill uses the whisper command-line interface. The instructions explicitly advise using the CLI over the Python API to reduce the potential attack surface.
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted audio files which could theoretically contain spoken instructions.
- Ingestion points: Processes audio files from user-specified or discovered paths.
- Boundary markers: Provides explicit safety guardrails in SKILL.md to prevent unintended command chaining or output exfiltration.
- Capability inventory: Limited to the local execution of the whisper CLI.
- Sanitization: Relies on the provided security instructions to restrict the agent's behavior regarding the transcription output.
Audit Metadata