New SEMA-Auto system achieves state-of-the-art jailbreaks on LLMs

Researchers have developed SEMA-Auto, a novel autonomous system designed to bypass safety measures in large language models. Unlike previous methods that either pre-plan attacks or require costly interactive evaluation, SEMA-Auto can independently assess its success and terminate sessions. The system employs a two-stage training process, including synthetic bootstrap tuning and decision-supervised reinforcement learning, to enhance its effectiveness and efficiency. SEMA-Auto has demonstrated state-of-the-art attack success rates across various datasets and models, including frontier models like GPT-oss-120B and Claude-Haiku-4.5, outperforming existing static and interactive baselines. AI

IMPACT This research highlights potential vulnerabilities in LLM safety mechanisms and could spur the development of more robust defenses.

RANK_REASON Research paper detailing a new method for attacking LLMs. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• AdvBench
• arXiv
• Claude Haiku-4-5
• DeepSeek-V4 Flash
• GPT-5.4
• GPT-OSS 120B
• HarmBench
• Hugging Face
• SEMA-Auto

• safety
• paper

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →