BEAST AI needs just a minute of GPU time to make an LLM fly off the rails

Researchers at the University of Maryland introduced **BEAST**, a fast beam-search adversarial prompting method that can jailbreak or induce hallucinations in LLMs in about a minute on a single GPU, achieving up to ~89% success on Vicuna-7B and potentially targeting public models that expose token probability scores. The technique can also strengthen membership inference attacks and craft more readable prompts suitable for social engineering, while stronger alignment (e.g., LLaMA-2) reduces—but does not eliminate—its effectiveness, underscoring the need for provable safety guarantees for deployed LLMs.