MemoryDojo: Benchmarking Poisoning Through LLM Agent Memory
All five memory systems tested were susceptible to poisoning, with substantial differences in the effects on later agent behavior.
MemoryDojo traces how malicious instructions encountered in one session are stored, retrieved and acted on in later sessions. The benchmark contains 155 multi-session, long-horizon trajectories across four domains. The authors also propose a memory-aware poison optimization process that increases attack efficacy, highlighting differences between memory architectures.
ArXiv · AI/CL/LG's note
MemoryDojo traces how malicious instructions encountered in one session are stored, retrieved and acted on in later sessions. The benchmark contains 155 multi-session, long-horizon trajectories across four domains. The authors also propose a memory-aware poison optimization process that increases attack efficacy, highlighting differences between memory architectures.
ArXiv · AI/CL/LG's note
score 5