Berkeley et al. propose a controllable-release prompt attack at USENIX Security 2026; large-model front-end filtering has a fundamental flaw
USENIX Security 2026|伯克利:一种 “可控释放提示词” 攻击曝光!大模型前置安全过滤器存在底层硬伤
Researchers from Berkeley and other institutions propose in a USENIX Security 2026 paper a controllable-release prompt attack that encodes jailbreak instructions into pseudo time-lock puzzles that lightweight guard models cannot solve but the main model can decrypt, bypassing large-model front-end input filtering. In black-box tests of the Gemini 2.5 Flash, DeepSeek Chat, xAI Grok, and Mistral Le Chat web versions, the timed-release variant achieved a 100% success rate against the first three, and 92% against Mistral. An evaluation of 14 open-source guard models on a total of 4000 samples showed that lightweight classifiers ranging from tens of M to 2B almost all failed, and DeepSeek-R1 could identify them but with a 3-70x increase in latency. The article's analysis holds that ChatGPT, Claude, and Meta products resisted the attack through output-side detection, and notes secondary risks such as leakage of intermediate chain-of-thought content and extraction of copyrighted children's books from Gemini (similarity up to 94%), recommending a shift to multi-layer defense in depth.
The paper empirically tests multiple commercial products and finds their front-end filtering can be bypassed with a high success rate; teams relying on lightweight guard models can use this to review their multi-layer defenses.