AI
BlindBias cuts black-box LLM attack API calls by 92.9% in Adobe, Amazon and USC study
Researchers from Adobe, Amazon and the University of Southern California have introduced BlindBias, a controlled decoding attack that uses sampled text alone. The method targets black-box large language models without model weights or token probabilities, and its speculative verification step reduces API calls from 4,000 to 283 in reported testing.