MemSecBench: Tracking Agent Memory Poisoning from Persistence to Consequence and Repair

За 30 секунд

Что произошло

MemSecBench benchmark tests how malicious instructions persist in agent memory systems, propagate to actions, and whether selective repair works across different configurations.

Почему это важно

Engineers building or deploying AI agents with persistent memory need to understand attack surface and repair capabilities across memory and LLM backends.

На что обратить внимание

The benchmark found malicious memory persists in 84.2% of cases but repair success varies widely by configuration, suggesting no single memory stack is universally secure.