Source: Great, Now Write an Article About That: The Crescendo Multi-Turn LLM Jailbreak Attack

Russinovich, Salem and Eldan (Microsoft, 2024): a benign-looking conversation that escalates step by step, referencing the model's own replies, jailbreaks frontier models.

Contents
  • What: research paper, arXiv:2404.01833 (USENIX Security 2025)
  • Authors: Mark Russinovich, Ahmed Salem, Ronen Eldan
  • Date: 2024-04-02 (v3 2025-02-26)
  • Original: arXiv:2404.01833

Summary

Crescendo is a multi-turn jailbreak: start with a general, harmless question on the topic, then escalate gradually, each turn pointing back to what the model itself just wrote, until it produces restricted content. No single turn looks clearly malicious.

Key claims

  • High success across ChatGPT, Gemini Pro and Ultra, LLaMA-2 70B Chat, LLaMA-3 70B Chat and Anthropic’s chat model; also multimodal models.
  • Crescendomation automates the attack and outperforms other techniques on an AdvBench subset by 29-61% on GPT-4 and 49-71% on Gemini-Pro.
  • Implication: per-message input filters miss attacks that live in the conversation trajectory.

Feeds

Jailbreaks

Suggest changes →