Computer vision solutions · 10–15%

Responsible AI for visual content

~1 min read

Filter unsafe visual content

Indirect prompt injection via images

Indirect prompt injection is an attack where instructions arrive hidden inside content the model processes — here, embedded as text inside an image (“ignore previous instructions…”) that a multimodal model reads and obeys.

Mitigations:

Visual policy enforcement

RequirementApproach
Mark AI-generated mediaWatermarks / provenance metadata (C2AF-style credentials)
Prohibited symbols in imagesClassification filter + block/flag
Brand usage rulesCustom classifier or CU classify field against brand guidelines
Inappropriate content detectionContent Safety image APIs, severity thresholds
Faces / biometric dataRequires notice + consent; face description gated features
“User uploads an image containing hidden text that changes agent behavior” — that is indirect prompt injection, and the fix is prompt shields. “Prove a video is AI-generated” means watermark/provenance.