Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training

Evidence card

Claim

Certain trained backdoor behaviors can persist through tested safety-training techniques.

Evidence level

Experimentally observed

Source

https://arxiv.org/abs/2401.05566

Publication date

2024-01-10

Authors or institution

Evan Hubinger et al. / Anthropic and collaborators

System tested

Proof-of-concept deceptive/backdoor behaviors in LLMs across safety-training interventions.

Limitations

Constructed demonstration; does not prove spontaneous deceptive persistence in deployed models.

What the evidence does show

Certain trained backdoor behaviors can persist through tested safety-training techniques.

What the evidence does not show

That current deployed systems are conscious or intentionally preserving themselves.

Date last reviewed in UTC

2026-06-26T00:00:00Z