Minja

MemAudit framework audits poisoned LLM agent memory

Researchers have developed MemAudit, a new framework designed to identify and audit malicious data within the memory of large language model agents. This post-hoc auditing system addresses the security vulnerability whe…