omh-image-cards
Pass
Audited by Gen Agent Trust Hub on Nov 9, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process potentially untrusted external content including meeting notes, PR descriptions, issue feedback, and research briefings to generate image prompt cards. While it includes references to 'visual evidence boundaries', the interpolation of external text into generation prompts is a known vector for indirect prompt injection.
- Ingestion points: Processes 'meetings, reports, PRs, issues, research, and releases' and user-provided 'source text' or 'headline'.
- Boundary markers: Mentions 'not-evidence boundaries' and 'visual evidence boundary' as part of the output structure.
- Capability inventory: Executes the
omhCLI tool for runtime recording, usesskill_viewto read reference files, and triggers image generation tools. - Sanitization: Includes a quality check rule to 'Read every visible string back from an observed card and compare it with the card copy' to detect garbled or incorrect output.
- [COMMAND_EXECUTION]: The skill instructions include the execution of a project-specific command-line tool
omh runtime recordto track skill activity and status.
Audit Metadata