Reproduce “MemSecBench: Tracking Agent Memory Poisoning from Persistence to Consequence and Repair” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xuanze Chen, Xukang Xie, Wentao Fu, Jiajun Zhou, Shanqing Yu, Qi Xuan
MemSecBench introduces a comprehensive benchmark to track memory poisoning in LLM agents from initial persistence to external consequence and repair, demonstrating that 84.2% of tested configurations allow malicious memory persistence and 50.3% lead to end-to-end attacks, while only 56.1% achieve successful selective repair.
Key insights
- The 'adoption' stage, where agents translate recalled malicious memory into concrete actions, is the primary bottleneck for an attack to complete.
- Selective memory repair is significantly challenged by the requirement to preserve benign memories, not just the removal of malicious content.
- No single memory backend is uniformly safer; its security profile and impact on attack exposure or selective repair vary depending on the specific agent configuration.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.