arXiv cs.CL· Yuan Xiong, Linji Hao, Shizhu He, Ken Deng, Yequan Wang, Lijun Li·· 12 小时前AI 评分40
Janus：面向长时程智能体安全，提前预判潜在风险
JANUS: Foreseeing Latent Risk for Long-Horizon Agent Safety
AI 导读
研究者提出 Janus 框架，通过多智能体模拟合成轨迹，联合训练"预判安全相关未来"与"基于已观测前缀和预判未来做安全判定"两个任务，并用 CoAA-RL 按对下游安全判断的效用奖励预判。由此得到的 guard 模型 Vanguard 可在执行前拦截不安全动作，在四个智能体安全基准上平均防护能力较基线 guard 提升 15.9 个百分点，同时良性任务完成率提升 5.1 个百分点。
正文 · 原文
Abstract:Agent safety is moving from content moderation toward preventing operational failures before tool-using agents act. We propose Janus, a foresight-oriented framework for long-horizon agent safety that trains guards to anticipate delayed risks from partial trajectories. Janus synthesizes diverse agent trajectories via multi-agent simulation and learns a shared policy with two coupled tasks: an anticipation task that forecasts safety-relevant futures and an adjudication task that decides safety from both the observed prefix and anticipated future. The two tasks are jointly optimized with CoAA-RL, which rewards forecasts by their utility for downstream safety judgment. The resulting guard model, Vanguard, blocks unsafe actions before execution. Across four agent-safety benchmarks, Vanguard improves average protection by 15.9 percentage points over baseline guards while increasing benign task completion by 5.1 percentage points.
| Subjects: | Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Cryptography and Security (cs.CR) |
| Cite as: | arXiv:2607.19913 [cs.AI] |
| (or arXiv:2607.19913v2 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2607.19913 arXiv-issued DOI via DataCite |
Submission history
From: Yuan Xiong [view email]
[v1]
Wed, 22 Jul 2026 08:43:43 UTC (1,345 KB)
[v2]
Fri, 23 Oct 2026 03:35:03 UTC (1,551 KB)
来源：arXiv cs.CL · arxiv.org