CoDA enforces security policies inside diffusion LLM decoding, 0% injection success
nandofioretto · x · 2026-10-28
arXiv paper 2610.15177 introduces CoDA (Policy-Constrained Diffusion Agents): exploiting the iterative, sequence-wide denoising of diffusion LLMs, it projects token distributions onto policy-compliant sets at every denoising step, enforcing task-level security policies without retraining. With oracle policies, adaptive prompt-injection success drops to 0% across 4 AgentDojo suites and 2 diffusion model families. Task policies are set from trusted instructions before untrusted content is seen, and protected fields can be reconsidered throughout decoding.
More from Safety
- Socket resurfaces mini shai-hulud lore: package hit via version range without touching latest — banteg · 2026-10-28
- eMudhra intermediate CA key compromise flagged P1 by Mozilla amid private comms — kingslyj · 2026-10-28
- Security researchers weigh whether AI-driven economic disruption will fuel global cybercrime — joshua_saxe · 2026-10-28
- AI policy researcher Janet Egan joins IFP as director of emerging tech — RishiBommasani · 2026-10-28
- HEIF Heist: image parser bugs could dump OpenAI repos and RCE Meta's suite — Miles_Brundage · 2026-10-28
- AI cyber-wreckage noise far outstrips actual damage so far, argues new analysis — timigod · 2026-10-28