Re-Mask and Redirect: Exploiting Denoising Irreversibility in Diffusion Language Models

Build a red-teaming tool that exploits denoising irreversibility in diffusion LMs. This helps safety researchers harden their models against prompt-injection and adversarial re-masking.

Suggested repo: diff-hack

"Bypass safety filters in diffusion LMs with a two-step re-masking attack."

Estimated effort: 50h