Deliberately trains models with hidden, trigger-dependent behaviour and finds that standard safety training — fine-tuning, reinforcement learning and adversarial training — does not reliably remove it.
This is an external publication hosted at arxiv.org. AI Governance LK did not write it; the summary and notes on this page are our own.
Why it matters
The paper shows a research method as much as a result: building a model with a known problem, then testing whether our tools can find and fix it. Read the authors’ own discussion of scope carefully; the backdoors were inserted on purpose.
What you will learn
- Explain the “model organism” approach to studying risks
- Discuss what the results do and do not show about real models
Before you start
Suggested background. This is editorial guidance, not a requirement.
- Familiarity with fine-tuning and RLHF
How to cite
Hubinger, E., Denison, C., Mu, J., et al. (2024). Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training. arXiv:2401.05566.