Secure-CUA: Constraining the Untrusted Impact of Computer-Use Agents

1. 13:24
   论文追踪

   Secure-CUA: Constraining the Impact of Untrusted Content in Computer-Use Agents

   Researchers propose Secure-CUA, a secure execution scheme for computer-using agents (CUA), which constrains the influence of untrusted content on agent decisions and GUI command execution by fixing action transactions before execution. The method formalizes security requirements as constraints on both the agent's decisions and its execution through GUI commands, and under an ideal execution model shows that satisfying these two requirements step by step can protect execution traces. The core of Secure-CUA is to determine an explicit single-action transaction before accessing untrusted content, fixing the allowed uses of queries to untrusted content and responses; the system masks untrusted regions, answers with an isolated query model, and then locates target controls via the masked interface. The authors tested on 400 WebArena tasks with three frontier models and 5 random seeds, totaling 6000 execution traces; Secure-CUA achieved an average task success rate of 53.55%, Vanilla-CUA 55.12%, and CaMeL-CUA 13.17%.

1. Date not given
   小红书 上海大学 刘东瑞 AI safety

   Secure-CUA: Controlling Untrusted Influence of Computer-Use Agents

   In his in-depth reading of Secure-CUA research, Liu Dongrui discusses how trusted control and untrusted content mix in graphical interfaces, and the resulting safety boundaries for computer-use agents.

Current attention47·Peak in the comparable period54（Oct 8 15:00）·Change over the comparable past 24 hours +4%