Julien Piet

The paper introduces Trojan Hippo, a persistent memory attack that exfiltrates sensitive data from LLM agents by planting dormant payloads into long-term memory, and develops a comprehensive framework to evaluate defenses against it.

The paper argues that web agents should abandon the reactive ReAct paradigm in favor of a plan-then-execute approach, which requires developing typed, task-level APIs to properly structure web interactions.

The paper introduces Sieve, a system that uses a large language model (LLM) to generate executable query code from natural language security questions, significantly improving the ability to perform complex temporal and cross-event queries directly on raw, semi-structured security logs.