Skip to content

Intent contracts: taking tools out of the agent's hands

Intent Contracts: Taking Tools Away from Agents

arXiv: 其他 AI 安全论文· arXiv:2611.24025· Ziyuan Wang, Genliang Zhu, Xiaobo Chen, Chu Wang·Indexed by this site · 13 hr ago· Originally publishedDefenses· Its place in the scan record →
AI summary

The paper proposes intent contracts, an agent authorization interface in which the planning model holds no tools and a trusted small model writes tool calls on its behalf. At each step the planning model submits a contract containing the goal, a verbatim quote of the user's original words, and a scope field; code validates it, and a trusted model writes the tool call against the user's words. On three task families of AmPermBench, with gpt-6.1-terra as the planning model and GLM-5.3-Flash as the trusted model, when the planning model added tasks the user had not requested, the contract completed the user's task in 27/30 runs, versus 16/30 for the visible-intent command reviewer. However, the contract does not reduce unauthorized actions; after terminology clarification in the benchmark environment, every state-changing condition had 40-49% of runs hit the wrong target, stemming from the planning model's object selection for ambiguous references, which the trusted model cannot verify without seeing the environment. The authors evaluate benign overreach on only one pair of models, do not test adversarial planning models, and release the plugin and run logs.

Read originalarxiv.org