creating-replay-vision-scanners
Pass
Audited by Gen Agent Trust Hub on Oct 27, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill describes a process for creating automated scanners that apply LLM prompts to session recordings, which are untrusted external data sources.
- Ingestion points: Session recordings matched by the
RecordingsQuerydefined inSKILL.mdare processed by the agent using vision models. - Boundary markers: The provided instructions do not include specific recommendations for using delimiters or 'ignore' instructions to isolate the scanner prompts from potential malicious instructions embedded within the recording data.
- Capability inventory: The skill utilizes tools such as
vision-scanners-create,vision-scanners-update, andvision-scanners-backfills-createto persist these automated scanning routines. - Sanitization: There is no mention of sanitizing or escaping the content retrieved from recordings (e.g., metadata, events, or console logs) before it is passed to models like
gemini-3-flash-preview.
Audit Metadata