designing-replay-vision-scanners
Pass
Audited by Gen Agent Trust Hub on Oct 10, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill designs prompts intended to process untrusted session recording data. While it includes defensive guidance, the workflow involves fetching external data and potentially modifying prompts based on that data.\n
- Ingestion points: The skill utilizes
vision-scanners-scan-sessioninSKILL.mdto fetch session recordings for calibration (Step 6).\n - Boundary markers: The skill explicitly instructs to include a section defining recording text as evidence rather than instructions (Step 4, item 8).\n
- Capability inventory: The
vision-scanners-prompt-suggestions-applytool allows updating configuration based on session observations.\n - Sanitization: The skill mandates privacy rules (Step 4, item 7) to paraphrase content and exclude sensitive identifiers.
Audit Metadata