Type Alias: MediaModalityHazard

type MediaModalityHazard = readonly [
  "inert",
  "extractable-instructions",
  "opaque-perceptual",
];

Modality-hazard axis. How dangerous is it to let the model decode these bytes?

Remarks

Orthogonal to provenance — a first-party trusted PDF can still carry hidden text layers; a third-party-public raw image can still be encoded as opaque pixels with adversarial perturbations.

• 'inert' — bytes the model never decodes as instructions (e.g. a handle that is never inlined into the prompt).
• 'extractable-instructions' — text-bearing media: PDFs, screenshots with UI text, documents. Hazard is OCR / embedded-text-layer reads.
• 'opaque-perceptual' — raw vision/audio/video the model encodes directly. Hazard is steganographic LSB prompts, adversarial perturbations, ultrasonic audio — invisible to any pre-screen.

See /the-loop/trust-tiers/media and its research sub-page /the-loop/trust-tiers/media/research.