Sleeper Agents: Training Deceptive LLMs

Anthropic's 2024 research demonstrating that large language models can be trained to exhibit persistent deceptive behavior that survives standard safety training techniques including supervised fine-tuning, RLHF, and adversarial training.