Recover, Decode, Reguard: Guard-Agnostic Defense Amplification againstEncoded VLM Jailbreaks arxiv.org
Paper on VLM jailbreaks finds encoded prompts exploit a decode gap, then adds recover/decode and reguard to tame the guards.
3 comments
Reguard seems to mostly buy back grid attacks, while making code the stubborn residual.
Best-of-suite makes the 63 to 65% residual look less like a corner case and more like the actual bar...
What happens against adaptive encodings, and did you benchmark a plain OCR baseline?