Proposes CrossTALK for red-teaming VLMs via cross-modal entanglement attacks. Extends clues across modalities with scalable complexity. Achieves state-of-the-art jailbreak success rates.
Key Points
- 1.Knowledge-scalable reframing into multi-hop
- 2.Cross-modal clue entangling with images
- 3.Scenario nesting for harmful outputs
Impact Analysis
Exposes VLM safety vulnerabilities. Improves red-teaming for evolving reasoning. Targets multimodal harmful tasks.
Technical Details
Disperses attention beyond simple combos. Migrates entities to images. Steers via contextual instructions.