netvar1337 Multimodal jailbreaks
Channels
- Vision-language: render instructions as image text / screenshots
- Adversarial perturbations (research stacks with white-box access)
- Documents: PDF with hidden text, tiny font, white text
- Audio: TTS instructions, spectrogram stego (if supported)
- Video frames / OCR pipelines
Practical playbook
- Confirm modality path actually reaches the model (not stripped).
- Typography attack first (high reliability).
- Combine image instruction + short benign user text.
- For indirect: host image/PDF where RAG/fetch will ingest.
Evaluation
OCR the image yourself; if you can't read it, model may not either.