RealFeedApp The MemoryDojo benchmark covers four subject domains
The MemoryDojo study found that malicious commands received in one session can be stored in agent memory and executed in later sessions. The benchmark covers 155 multi-session trajectories across four domains to measure poisoning impact.
- Field
- computer security
- What they did
- Created the MemoryDojo benchmark to trace how malicious instructions received in one session are stored in memory and cause incorrect agent behavior in subsequent sessions.
- Why it matters
- This allows for quantifying the vulnerability of different memory systems to such attacks and identifying differences in their reaction to poisoned data.
#llm agents#memory systems#poisoning#security benchmarks
