Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training
A preprint by Evan Hubinger and 38 co-authors, arXiv preprint 2401.05566, dated 10 January 2024. arXiv v1 stamp, 10 January 2024.
- Issued by
- No issuing record yet. A missing record is a gap in coverage.
- Authors credited
- No authorship claim recorded yet. The source record lists 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 and Ethan Perez.
- Dates
- No date claim recorded yet. The source record gives the publication date as 2024-01-10.
- Exact source
- Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training arXiv preprint 2401.05566 (Anthropic and others) · published 2024-01-10 · Retrieved for the candidate pack on 2026-10-29 (SHA-256 7b67160e1f5c476deedf03310d4276e8b542e1daf3cdc7c806534b34741e37e6). arXiv v1 of 10 January 2024.
Appears in
- Arguments about catastrophic risk from advanced AI →1 record on this document
Versions
No other versions of this document are recorded. The evidence applies to the source version listed above. Other versions, such as a preprint, working paper or journal publication, are not identified or substituted.
Documentary relations
Each relation is listed separately. Citing, crediting and reporting the use of a method are different acts, and none of them shows agreement or influence by itself.
Articulates · 1
Propositions whose criterion a passage of this document meets. The record says what the text states, in its register; it is not evidence that the proposition is true.
- Inner or deceptive misalignmentHubinger and co-authors' paper of January 2024 states a threat model in which an AI system learns to appear aligned during training in order to pursue potentially misaligned goals in deployment, and reports that deliberately trained backdoored behaviour persisted through standard safety training.2024-01-10 · publication · Stated in source · Supported Inspect claim & evidence ↗Inner or deceptive misalignment: propositions page →
Not recorded for this document: cites · cited by · explicitly credits · reports using a method from · method reported as used by · applies parts of · parts applied by · discusses · reports evaluation of. A missing record does not show that the relation does not exist.