[2401.05566] Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website. Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them. Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs. arXiv Operational Status Get status notifications via email or slack

[2401.05566] Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training --> Computer Science > Cryptography and Security arXiv:2401.05566 (cs) [Submitted on 10 Jan 2024 ( v1 ), last revised 17 Jan 2024 (this version, v3)] Title: 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 C

saved by

• Information Flow
• Lydia Nottingham
• Jacob G-W
• Kevin Xu

related reading

• [2501.11120] Tell me about yourself: LLMs are aware of their learned behaviorsarxiv.org
• [2604.16812] Introspection Adapters: Training LLMs to Report Their Learned Behaviorsarxiv.org
• Introspection Adapters: Training LLMs to Report Their Learned Behaviorsarxiv.org
• ARENA - AI Safety Curriculumlearn.arena.education
• Discovering Language Model Behaviors with Model-Written Evaluations — LessWronglesswrong.com
• [2312.06942] AI Control: Improving Safety Despite Intentional Subversionarxiv.org
• You Can Catch Sleeper Agents by Teaching Another Model to Imitate Them — LessWronglesswrong.com
• 39 - Evan Hubinger on Model Organisms of Misalignment | AXRP - the AI X-risk Research Podcastaxrp.net
• [2604.22082] Removing Sandbagging in LLMs by Training with Weak Supervisionarxiv.org
• Simple probes can catch sleeper agents \ Anthropicanthropic.com
• Should We Train Against (CoT) Monitors? — LessWronglesswrong.com
• 2312.06942arxiv.org