Source note3 min read
SHE: Trajectory-driven Safety Harness Evolution for LLM Agents
Source
arXiv
arxiv.org/abs/2608.09885v1
Source
- Source name: arXiv
- Source kind: paper
- Authors / organization: Wanying Qu, Qinghua Mao, Yu Li, Jiyao Liu, Xin Zhang, Dadi Guo, Yanxu Zhu, Qingyu Liu, Leitao Yuan, Xi Lin, Shanfeng Zhu, Yanwei Fu, Jing Shao, Xia Hu, Dongrui Liu
- Venue: arXiv: cs.AI
- Published: 2026-08-10
- Collected: 2026-08-11
- URL: http://arxiv.org/abs/2608.09885v1
- PDF: https://arxiv.org/pdf/2608.09885v1
- DOI: UNKNOWN
- Registry group: academic_apis_and_preprints
- Evidence role: canonical_academic_index
- Language: en
- Country / region: UNKNOWN
Plain-Language Brief
Researchers propose a framework called Safety Harness Evolution (SHE) to improve the safety of large language model agents. SHE learns evolving safe boundaries from rollout trajectories, allowing for localized harness evolution. It decomposes the harness into four components with clear safety responsibilities and introduces an attribution-guided evolution loop.
What the Document Says
- The text discusses a framework called Safety Harness Evolution (SHE).
- SHE is designed to improve the safety of large language model agents.
- SHE learns evolving safe boundaries from rollout trajectories.
- SHE decomposes the harness into four artifacts: System Prompt, Rule Bank, Safety Memory, and Tool Policy.
- Each artifact has explicit safety responsibilities.
- The decomposition provides clear functional boundaries for localized evolution.
- SHE introduces an attribution-guided evolution loop.
- This loop converts trajectory failures into structured diagnoses.
- It learns artifact-specific boundary refinements.
- Evolved harnesses are selected through safety-utility validation.
- Experiments show that SHE enhances safety and improves utility compared to static SafeHarness.
Key Concepts Mentioned
- Safety Harness Evolution (SHE)
- Large Language Model (LLM) agents
- Rollout trajectories
- Decomposition of harness components
- Attribution-guided evolution loop
- Artifact-specific boundary refinements
Claims Made by the Document
- SHE improves the safety of large language model agents.
- SHE achieves a 3.1x ASR reduction compared with static SafeHarness.
Evidence Used
- The text describes SHE as a framework to improve the safety of LLM agents.
- The text mentions an experiment demonstrating a 3.1x ASR reduction for SHE compared to static SafeHarness.
Methods / Approach
- Learning evolving safe boundaries from rollout trajectories
- Decomposing the harness into four artifacts with explicit safety responsibilities
- Introducing an attribution-guided evolution loop
Limitations Stated by the Authors
The text does not mention any limitations as stated by the authors.
Limitations Noted by the Agent
Agent interpretation — not from the source.
The agent's limitations include interpreting the impact of the proposed method on real-world applications, evaluating its efficiency and scalability, and understanding its potential unintended consequences.
Open Questions
- What is the ASR metric mentioned in the text?
- How does SHE generalize to other types of agents or risks not covered by Agent-SafetyBench and AgentHarm benchmark?
Notes for Future Q&A
May include interpretation. Labeled where it does.
- ASR likely refers to Automatic Speech Recognition.
- The text mentions that the evolved harness generalizes, but it does not provide details on how extensively.
Raw Abstract / Excerpt
The safety of large language model (LLM) agents depends not only on model weights but also on the agent harness that manages context, memory, tools, permissions, and runtime control. Existing safety mechanisms often treat the harness as a fixed deployment artifact, limiting their ability to evolve with emerging risks. Moreover, coupled functions across harness components obscure safety responsibility attribution, making localized evolution difficult. We propose Safety Harness Evolution (SHE), a framework that learns evolving safe boundaries from rollout trajectories. SHE decomposes the harness into four artifacts with explicit safety responsibilities, including the System Prompt, Rule Bank, Safety Memory, and Tool Policy, defining clear functional boundaries for localized evolution. Based on this decomposition, SHE introduces an attribution-guided evolution loop that converts trajectory failures into structured diagnoses, learns artifact-specific boundary refinements, and selects evolved harnesses through safety-utility validation. Experiments on Agent-SafetyBench demonstrate that SHE effectively enhances safety through harness evolution, achieving a 3.1x ASR reduction compared with static SafeHarness, while also improving benign utility. The evolved harness further generalizes to unseen risks on the held-out AgentHarm benchmark and transfers across agent models without additional evolution.
