Cardei et al.
Diffusion language models enforce task-level security against prompt injection
CoDA projects token distributions onto policy-compliant sets at each denoising step, eliminating successful attacks under oracle policies.
Big Tech
Michael Cardei · Anjun Hu · Chenrui Ma · Francesco Pinto · Chawin Sitawarin · Chulin Xie · +1 more
University of Virginia · University of Oxford · Google DeepMind
Research Digest··3 min read
Thread:Agent Security & Attacks
The authors introduce CoDA, a training-free framework that embeds task-level security policies directly into the iterative denoising process of diffusion language models.
Why this paper
From Google DeepMind and 2 others · Part of Agent Security & Attacks, now 89 papers
In one line
CoDA prevents prompt injection by enforcing task-level security policies within diffusion language model denoising.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§