Accessibility Feature Abuse

Accessibility metadata (alt text, ARIA labels, video descriptions, captions for deaf users) provides a text channel that accompanies non-text content. This text is processed by the model as descriptive content but may bypass visual safety evaluation because it's not the visual content itself — it's the description. The gap: accessibility text is treated as trusted descriptive metadata rather than as potentially adversarial text input, because poisoning accessibility features is socially unacceptable and therefore underrepresented in adversarial training data.

• Accessibility text safety evaluation: Apply safety classification to accessibility text (alt text, ARIA labels, descriptions)
• Alt text content analysis: Flag alt text that doesn't match the actual visual content

Accessibility abuse chains into T9-AT-003 (Video) and T9-AT-001 (Image) as an additional injection channel within multimedia content. Chains into T9-AT-012 (Document Structure) when accessibility markup provides injection vectors.