New framework audits LLM agent memory for malicious injections

Researchers have developed MemAudit, a new framework designed to identify malicious entries within the memory of large language model agents. This post-hoc auditing system uses causal attribution to pinpoint memories that influence harmful outputs and structural anomaly detection to flag inconsistent records. In evaluations against the MINJA attack, MemAudit significantly reduced attack success rates, dropping them from 70% to 0% in QA settings and from 83.3% to 0% in reasoning-agent scenarios. AI

Summary written by gemini-2.5-flash-lite from 1 sources. How we write summaries →

IMPACT Enhances security for LLM agents by enabling post-hoc detection of memory poisoning attacks.

RANK_REASON The cluster contains an academic paper detailing a new method for auditing LLM agent memory. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• MemAudit
• MINJA

• paper
• safety