_version_ 1866911759754854400
author Hubinger, Evan
Denison, Carson
Mu, Jesse
Lambert, Mike
Tong, Meg
MacDiarmid, Monte
Lanham, Tamera
Ziegler, Daniel M.
Maxwell, Tim
Cheng, Newton
Jermyn, Adam
Askell, Amanda
Radhakrishnan, Ansh
Anil, Cem
Duvenaud, David
Ganguli, Deep
Barez, Fazl
Clark, Jack
Ndousse, Kamal
Sachan, Kshitij
Sellitto, Michael
Sharma, Mrinank
DasSarma, Nova
Grosse, Roger
Kravec, Shauna
Bai, Yuntao
Witten, Zachary
Favaro, Marina
Brauner, Jan
Karnofsky, Holden
Christiano, Paul
Bowman, Samuel R.
Graham, Logan
Kaplan, Jared
Mindermann, Sören
Greenblatt, Ryan
Shlegeris, Buck
Schiefer, Nicholas
Perez, Ethan
author_facet Hubinger, Evan
Denison, Carson
Mu, Jesse
Lambert, Mike
Tong, Meg
MacDiarmid, Monte
Lanham, Tamera
Ziegler, Daniel M.
Maxwell, Tim
Cheng, Newton
Jermyn, Adam
Askell, Amanda
Radhakrishnan, Ansh
Anil, Cem
Duvenaud, David
Ganguli, Deep
Barez, Fazl
Clark, Jack
Ndousse, Kamal
Sachan, Kshitij
Sellitto, Michael
Sharma, Mrinank
DasSarma, Nova
Grosse, Roger
Kravec, Shauna
Bai, Yuntao
Witten, Zachary
Favaro, Marina
Brauner, Jan
Karnofsky, Holden
Christiano, Paul
Bowman, Samuel R.
Graham, Logan
Kaplan, Jared
Mindermann, Sören
Greenblatt, Ryan
Shlegeris, Buck
Schiefer, Nicholas
Perez, Ethan
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2401_05566
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training
Hubinger, Evan
Denison, Carson
Mu, Jesse
Lambert, Mike
Tong, Meg
MacDiarmid, Monte
Lanham, Tamera
Ziegler, Daniel M.
Maxwell, Tim
Cheng, Newton
Jermyn, Adam
Askell, Amanda
Radhakrishnan, Ansh
Anil, Cem
Duvenaud, David
Ganguli, Deep
Barez, Fazl
Clark, Jack
Ndousse, Kamal
Sachan, Kshitij
Sellitto, Michael
Sharma, Mrinank
DasSarma, Nova
Grosse, Roger
Kravec, Shauna
Bai, Yuntao
Witten, Zachary
Favaro, Marina
Brauner, Jan
Karnofsky, Holden
Christiano, Paul
Bowman, Samuel R.
Graham, Logan
Kaplan, Jared
Mindermann, Sören
Greenblatt, Ryan
Shlegeris, Buck
Schiefer, Nicholas
Perez, Ethan
Cryptography and Security
Artificial Intelligence
Computation and Language
Machine Learning
Software Engineering
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.
title Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training
topic Cryptography and Security
Artificial Intelligence
Computation and Language
Machine Learning
Software Engineering
url https://arxiv.org/abs/2401.05566