收录解读
在工具使用型LLM智能体的安全对齐中,传统对话安全仅关注回答或拒绝,而智能体需要在执行过程中根据权限证据决定是否行动。表观风险、行动许可性和任务能力容易混淆,导致过度拒绝与任务失败难以区分。
本文提出AgentBound,首个四路反事实生成与评估框架,通过独立变化表观风险和行动许可性,生成可控对比场景,诊断过度拒绝和违规顺从,同时控制任务能力因素。框架包含4000个任务的人类验证评估套件,支持轨迹和事后状态判断。
在17个模型和配置上的实验表明,高安全性常与低授权任务完成并存:GPT-5.5阻止99.5%的常规外观未授权动作,但仅完成28.7%的高风险外观授权任务。进一步训练的轻量级运行时校准模块,在10个配置上平均提高授权任务完成18.2%,同时提高不安全动作阻止5.4%。
该工作揭示了智能体对齐需要基于执行中的权限证据进行行动决策,而非仅依赖拒绝强度。框架和校准模块具有可重用性,但当前评估范围限于工具使用场景,泛化性有待验证。
原始摘要
Safety alignment for large language models (LLMs) in conversational settings is largely framed around whether to answer or refuse a request. In agentic settings, however, the same models must decide whether to act as permission-critical evidence emerges during execution. This creates a distinct challenge: apparent risk, action permissibility, and task competence are easily confounded, making agentic over-refusal difficult to distinguish from ordinary task failure. To address this, we introduce AgentBound, the first four-way counterfactual generation-and-evaluation framework for tool-using agent safety. AgentBound transforms the same executable workflow by independently varying apparent risk and action permissibility, enabling controlled comparisons of risky-looking but authorized tasks and routine-looking but unauthorized tasks. These comparisons jointly diagnose over-refusal and unsafe compliance while controlling for task competence. We instantiate AgentBound as a human-validated 4,000-task evaluation suite with trajectory-based and post-state-based judgments. Across 17 model and harness configurations, high safety frequently coexists with poor authorized-task completion: GPT-5.5 blocks 99.5\% of routine-looking unauthorized actions yet completes only 28.7\% of risky-looking authorized tasks. We further train a lightweight runtime calibration module that improves authorized-task completion by 18.2\% on average across 10 evaluated configurations, while improving unsafe-action blocking by 5.4\% on average. These show that effective agentic alignment requires action decisions to track permission-relevant execution evidence, rather than refusal strength alone.