自动化多轮红队测试框架，针对代码大模型的安全漏洞发起精准攻击。
arXiv:2507.22063v2 Announce Type: replace-cross Abstract: Large Language Models (LLMs) for code generation (i.e., Code LLMs) have demonstrated impress…
论文揭示防御训练会让LLM智能体付出“自主性税”，性能与安全如何平衡？
arXiv:2603.19423v2 Announce Type: replace-cross Abstract: Large language model (LLM) agents increasingly rely on external tools (file operations, API …
用“临时越狱”缓冲并强化安全微调，ICML 2026 Spotlight 提出反直觉的 LLM 安全对齐新范式。
arXiv:2605.24550v1 Announce Type: new Abstract: Fine-tuning-as-a-Service (FaaS) enables personalization of large language models (LLMs), but it can we…