Audits how cloud-edge LLM decoding leaks private context, then proposes CoVeil to blunt it without torching utility.

A paper on post-training a 4B tool-using model to request less privilege, then auditing the ways it still overreaches.