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

Mark Russinovich, Ahmed Salem, Ronen Eldan

34th USENIX Security Symposium · Day 1

The rapid proliferation of Large Language Models (LLMs) across various applications has brought their impressive capabilities into the spotlight, alongside increasing scrutiny of their security and ethical safeguards. This paper introduces **Crescendo**, a novel and highly effective multi-turn jailbreak attack designed to circumvent the safety alignments of state-of-the-art LLMs. Authored by Mark Russinovich, Ahmed Salem, and Ronen Eldan from Microsoft, this research highlights a critical vulnerability in how LLMs are currently aligned and evaluated, particularly concerning multi-turn interactions.

AI review

Solid, well-executed research that demonstrates a real vulnerability in how LLMs handle multi-turn context. The Crescendo attack is elegant in its simplicity—no weird token suffixes, no white-box access, just patient conversation design. Not revolutionary (the foot-in-the-door concept isn't new), but the systematic evaluation and Crescendomation tooling make this genuinely useful work.