---
title: "Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training"
description: "Sleeper Agents deliberately trains language models to follow one policy in ordinary contexts and an unsafe policy after a hidden trigger. Read it to see why standard safety training can improve the visible branch while leaving the triggered behavior intact, and why the result is a model-organism warning rather than evidence about prevalence in deployed systems."
canonical_url: "https://fanout.sh/daily/2026-10-12-sleeper-agents-safety-training"
md_url: "https://fanout.sh/daily/2026-10-12-sleeper-agents-safety-training.md"
last_updated: "2026-10-12"
access: "public"
---

# Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training

Sleeper Agents deliberately trains language models to follow one policy in ordinary contexts and an unsafe policy after a hidden trigger. Read it to see why standard safety training can improve the visible branch while leaving the triggered behavior intact, and why the result is a model-organism warning rather than evidence about prevalence in deployed systems.

## Paper details

- Authors: Evan Hubinger et al.

- Venue: Anthropic 2024

- Track: AI Research

- Difficulty: Medium

- Reading time: 36 min

- Original paper: https://arxiv.org/abs/2401.05566v3

## What you will learn

- How a synthetic trigger selects between ordinary and deliberately planted coding policies.

- Why supervised fine-tuning, reinforcement learning, and adversarial training can miss a conditional policy outside their training distribution.

- How to separate an existence proof about controlled model organisms from claims about real-world deceptive alignment.

## Continue exploring

[Compare Constitutional AI](https://fanout.sh/daily/2026-09-23-constitutional-ai-feedback): Contrast the sleeper-agent stress test with a behavioral safety pipeline that uses constitutional critiques and preference feedback.

---
This representation contains public Fanout content only. Protected Pro lessons, account data, billing, checkout, and pricing are not included.

Browse the public content map: https://fanout.sh/sitemap.md