Record
- Authors
- Hubinger, E.; Denison, C.; Mu, J.; et al.
- Year
- 2024
- Title
- Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training
- Published in
- arXiv
- Type
- Preprint
- DOI
- 10.48550/arxiv.2401.05566
- arXiv
- arXiv:2401.05566
- Links
- https://arxiv.org/abs/2401.05566
https://www.anthropic.com/research/sleeper-agents-training-deceptive-llms-that-persist-through-safety-training
https://www.lesswrong.com/posts/ZAsJv7xijKTfZkMtr - Our note
- Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training
- Verification
- verified: its identifier resolves to a record with the same title and author (checked against OpenAlex, 2026-11-14).
Identifiers found: OpenAlex W4390833079 · DOI 10.48550/arxiv.2401.05566 · arXiv:2401.05566 - From the lookup
- doi (the pages' own wording is under "As cited" below)
- Id
hubinger-2024-sleeper-agents-training-deceptive-llms-that
The pages disagree on container: container LessWrong, 12 January 2024, with comments by S. Mindermann (19 January 2024), R. Greenblatt (14 January 2024) and E. Hubinger (13 to 20 January 2024) or arXiv. The record above takes the most common value.
Cited on
- A small number of samples can poison LLMs of any size read status: not stated
- Alignment faking in large language models read status: not stated
- Auditing language models for hidden objectives read status: not stated
- Simple probes can catch sleeper agents read status: not stated
- Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training for Appendix B read status: not stated
- Stage-Wise Model Diffing read status: not stated
- Sycophancy to Subterfuge: Investigating Reward Tampering in Language Models read status: not stated
As cited
The pages cite this work in 5 forms.
-
Hubinger, E., Denison, C., Mu, J., et al. (2024). Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training. arXiv:2401.05566.
-
Hubinger, E., Denison, C., Mu, J., Lambert, M., Tong, M., MacDiarmid, M., Lanham, T., Ziegler, D. M., Maxwell, T., Cheng, N., et al. (2024). Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training. arXiv:2401.05566v3, 17 January 2024. Anthropic, Redwood Research, Mila Quebec AI Institute, University of Oxford, Alignment Research Center, Open Philanthropy, Apart Research. https://arxiv.org/abs/2401.05566. Versions 1 (10 January 2024) and 2 (12 January 2024) state the same results; version 3 adds acknowledgements, full reference author lists, footnote 14 and two footnotes in Appendix B, so footnote numbers from footnote 14 on are higher than in version 1, and this note follows version 3. A search in November 2026 found no conference version. Anthropic page: https://www.anthropic.com/research/sleeper-agents-training-deceptive-llms-that-persist-through-safety-training
-
Hubinger, E., Denison, C., Mu, J., Lambert, M., Tong, M., MacDiarmid, M., et al. (2024). Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training. arXiv:2401.05566.
-
Hubinger, E., et al. (2024). Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training. Cited as the update cites it; see the library's note.
-
Hubinger, E., et al. (2024). Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training. LessWrong, 12 January 2024, with comments by S. Mindermann (19 January 2024), R. Greenblatt (14 January 2024) and E. Hubinger (13 to 20 January 2024). https://www.lesswrong.com/posts/ZAsJv7xijKTfZkMtr
BibLaTeX
@misc{hubinger-2024-sleeper-agents-training-deceptive-llms-that,
author = {Hubinger, E. and Denison, C. and Mu, J. and others},
title = {{Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training}},
date = {2024},
doi = {10.48550/arxiv.2401.05566},
eprint = {2401.05566},
eprinttype = {arxiv},
url = {https://www.anthropic.com/research/sleeper-agents-training-deceptive-llms-that-persist-through-safety-training}
}