Skip to content

Using information flow control to secure AI agents

Securing AI Agents with Information-Flow Control

arXiv· arXiv:2505.23643· NeurIPS 2026· Manuel Costa, Boris Köpf, Aashish Kolluri, Andrew Paverd, Mark Russinovich, Ahmed Salem, Shruti Tople, Lukas Wutschitz, Santiago Zanella-Béguelin·Indexed by this site · 22 days ago· Originally publishedDefenses
AI summary

The paper proposes using information flow control (IFC) to provide AI agents with security guarantees against vulnerabilities such as prompt injection. The authors built a formal model characterizing the security and expressiveness of agent planners, defined the classes of security properties enforceable by dynamic taint tracking, and designed a task taxonomy to measure the trade-off between security and utility for different planners. On this basis they proposed the planner Fides, which tracks confidentiality and integrity labels, enforces security policies deterministically, and introduces a new primitive for selectively hiding information; evaluation on the AgentDojo benchmark shows that the method can complete a relatively broad range of tasks with security guarantees. The accompanying tutorial is at https://github.com/microsoft/fides。

Related events and news

1 events · 3 news mentions

News that mentions this paper

微软构建信任感知控制应对间接提示注入

Read originalNeurIPS 2026