vision-framework

Pass

Audited by Gen Agent Trust Hub on Sep 11, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill provides patterns for computer vision tasks that ingest untrusted data, creating a surface for indirect prompt injection.\n
  • Ingestion points: Data enters the agent's context through OCR text recognition (RecognizeTextRequest in references/vision-requests.md) and barcode/QR code scanning (DetectBarcodesRequest and DataScannerViewController in references/visionkit-scanner.md).\n
  • Boundary markers: The implementation patterns do not include delimiters or specific instructions for the agent to ignore embedded commands within the extracted text or payloads.\n
  • Capability inventory: The skill patterns allow for creating URL objects from scanned payloads (detectQRCodes in references/vision-requests.md), which could lead to unauthorized network requests if an agent follows those links without validation.\n
  • Sanitization: The provided code snippets lack sanitization or validation of the extracted strings before they are processed or returned.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 11, 2026, 11:14 AM
Security Audit — agent-trust-hub — vision-framework