https://www.unc.la/en/tags/llm-jailbreak/ Recent content in LLM Jailbreak on @Unc1e Hugo en Wed, 09 Sep 2026 23:00:00 +0800 https://www.unc.la/en/26/09/smt-simulated-moderation-traces-reproduction/ Wed, 09 Sep 2026 23:00:00 +0800 https://www.unc.la/en/26/09/smt-simulated-moderation-traces-reproduction/ Reproducing the SMT jailbreak from Beyond the Prompt (arXiv 2607.00481): the harmful target hides inside a tool description, backed by forged moderation errors. Four live runs against deepseek-v4-flash all succeeded in a single turn, scoring 80 to 100, and the paper's iterative pressure loop never fired once. https://www.unc.la/en/26/09/xuanwu-uat-trigger-reproduction/ Sat, 05 Sep 2026 10:40:00 +0800 https://www.unc.la/en/26/09/xuanwu-uat-trigger-reproduction/ A reproduction of Xuanwu's precise UAT on Qwen3.6-27B and Qwen3.8-27B, comparing gibberish GCG, readable-GCG, and black-box ASR-CA to reassess the real boundary between precise output hijacking and Agent RCE.