omh-media-input
Pass
Audited by Gen Agent Trust Hub on Sep 25, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted external data from media sources including audio, video, YouTube links, and OCR outputs.
- Ingestion points: Media source locations, YouTube transcripts, OCR text from receipts/screenshots, and meeting recordings.
- Boundary markers: The skill explicitly utilizes
transcript_boundary/v1to separate supplied transcripts from missing ones and identifies 'not-evidence boundaries' to prevent hallucination. - Capability inventory: Involved in transcription, media summarization, and task status tracking via the
omhruntime harness. - Sanitization: Implements 'hallucination risk' gating and requires explicit permission boundaries to be defined for media sources.
- [COMMAND_EXECUTION]: The instructions contain a shell command template for recording workflow status using the local CLI tool
omh runtime record. This command is used for internal platform tracking and does not appear to interact with sensitive system files or non-whitelisted network domains.
Audit Metadata