跳到正文
arXiv cs.AI· Yujun Wang, Yilun Liu, Jinhe Bi, Krishna Kanth Nakka, Xuebing Zhou, Hinrich Sch{\"u}tze, Yunpu Ma·· 4 小时前AI 评分48

Evolution Under Fire:用自适应注入压力器进化智能体安全护栏

Evolution Under Fire: Evolving Agent Safety Harnesses with Adaptive Injection Stressors

AI 导读

研究提出 Evolution Under Fire 分阶段框架,将安全护栏进化与自适应压力器进化耦合,通过效用约束的 Pareto 调度树搜索从执行证据中进化出完整可执行护栏。

正文 · 原文

View PDF HTML (experimental)

Abstract:Tool-using language-model agents must act on information from untrusted emails, webpages, and documents. This creates a central security challenge: malicious instructions and legitimate task directives can arrive through the same interaction channel. Existing defenses are largely hand-designed and static, making further improvements dependent on repeated manual redesign. We introduce Evolution Under Fire, a staged framework that couples safety-harness evolution with adaptive stressor evolution. Utility-constrained, Pareto-scheduled tree search evolves complete executable safety harnesses from execution evidence while keeping the task model, benchmark, tools, and evaluator fixed. Adaptive stressor evolution then discovers utility-preserving attacks that guide a subsequent stage of harness evolution. On AgentDyn with GPT-4o-mini, an evolved harness reduces attack success from 50.4% to 3.0%, improves clean and attacked utility over the initial harness, and offers a stronger security-utility trade-off than reproduced defenses. On Agent-SafetyBench with Qwen2.5-7B-Instruct, harness evolution reduces attack success from 82.6% to 62.0% while retaining helpfulness and runtime validity. Evolved harnesses also transfer without further evolution to Qwen3.5-35B-A3B, reducing attack success from 32.1% to 0.9% while improving attacked-task utility.
Subjects: Cryptography and Security (cs.CR); Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
Cite as: arXiv:2611.22583 [cs.CR]
  (or arXiv:2611.22583v1 [cs.CR] for this version)
  https://doi.org/10.48550/arXiv.2611.22583

arXiv-issued DOI via DataCite (pending registration)

Submission history

From: Yujun Wang [view email]
[v1] Tue, 17 Nov 2026 11:07:31 UTC (992 KB)

来源:arXiv cs.AI · arxiv.org