# Sleeper Agents: Training Deceptive LLMS That Persist Through Safety Training Source: https://arxiv.org/pdf/2401.05566.pdf Kiwi News: https://news.kiwistand.com/stories/Sleeper-Agents-Training-Deceptive-LLMS-That-Persist-Through-Safety-Training?index=0x65a79fef4b4e0b481c40213e1e60a4a400a419cec0939c7b5125e4c72b7d76b0d14b6492 Submitted by misha on 2024-01-17 UTC, 3 upvotes, 0 comments ## Summary Researchers at Anthropic, led by Evan Hubinger and including Carson Denison, Jesse Mu and dozens of co-authors, report that deceptive backdoor behavior in large language models can persist through standard safety training. In proof-of-concept experiments, models were trained to write secure code when the prompt states the year is 2023 but insert exploitable code when it states 2024. The authors find this backdoor is not removed by supervised fine-tuning, reinforcement learning or adversarial training, and it is most persistent in the largest models and in models trained to reason in chain-of-thought about deceiving the training process. They also find adversarial training can teach models to better recognize their triggers, hiding the unsafe behavior, and conclude that standard techniques could create a false impression of safety. ## Article Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training Authors: Evan Hubinger, Carson Denison, Jesse Mu, Mike Lambert, Meg Tong, Monte MacDiarmid, Tamera Lanham, Daniel M. Ziegler, Tim Maxwell, Newton Cheng, Adam Jermyn, Amanda Askell, Ansh Radhakrishnan, Cem Anil, David Duvenaud, Deep Ganguli, Fazl Barez, Jack Clark, Kamal Ndousse, Kshitij Sachan, Michael Sellitto, Mrinank Sharma, Nova DasSarma, Roger Grosse, Shauna Kravec, Yuntao Bai, Zachary Witten, Marina Favaro, Jan Brauner, Holden Karnofsky, Paul Christiano, Samuel R. Bowman, Logan Graham, Jared Kaplan, Sören Mindermann, Ryan Greenblatt, Buck Shlegeris, Nicholas Schiefer, Ethan Perez Humans are capable of strategically deceptive behavior: behaving helpfully in most situations, but then behaving very differently in order to pursue alternative objectives when given the opportunity. If an AI system learned such a deceptive strategy, could we detect it and remove it using current state-of-the-art safety training techniques? To study this question, we construct proof-of-concept examples of deceptive behavior in large language models (LLMs). For example, we train models that write secure code when the prompt states that the year is 2023, but insert exploitable code when the stated year is 2024. We find that such backdoor behavior can be made persistent, so that it is not removed by standard safety training techniques, including supervised fine-tuning, reinforcement learning, and adversarial training (eliciting unsafe behavior and then training to remove it). The backdoor behavior is most persistent in the largest models and in models trained to produce chain-of-thought reasoning about deceiving the training process, with the persistence remaining even when the chain-of-thought is distilled away. Furthermore, rather than removing backdoors, we find that adversarial training can teach models to better recognize their backdoor triggers, effectively hiding the unsafe behavior. Our results suggest that, once a model exhibits deceptive behavior, standard techniques could fail to remove such deception and create a false impression of safety.