Sleeper Agents : Training deceptive LLMs that persist through safety training构建带条件后门的模型,并测试相关行为能否在安全训练后持续存在。核对原文署名 ↗