What's this about?
This paper shows that diffusion-based language models have their own distinct safety weak spots — analyzing them both as *targets* (things to attack) and as *adversaries* (things that attack), with mechanistic explanations of why.
What is a "diffusion LLM"?
Most LLMs generate text left-to-right, one token at a time. A diffusion LLM instead starts from noise and iteratively refines a whole block of text at once — a technique borrowed from image generators. Because the generation process is fundamentally different, its safety failure modes are different too, so jailbreak defenses built for normal LLMs may not transfer.
Why does it matter?
Diffusion LLMs are an emerging alternative architecture; showing they need their own safety analysis is a heads-up as more labs experiment with them.