AI safety researchers find method to block jailbreak attacks across text, image, and audio inputs

Researchers created a simple technique that detects harmful requests before a multimodal AI generates responses, reducing successful attacks by 97% across different input types. The technique works by checking internal patterns in the AI's processing without needing external tools or retraining on specific input types.

This matters because multimodal AI systems (ones that take text, images, and audio) are harder to keep safe than text-only versions — attackers can exploit the gaps between how the model handles different input types. The technique is simple enough that deployed systems could add it without major redesign, which means the safety problem might actually be solvable without expensive overhauls. The real test is whether this holds up when facing adversaries actively trying to break it, not just existing attack benchmarks.