SHE: Trajectory-driven Safety Harness Evolution for LLM Agents

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) for improving the safety of large language model agents. SHE focuses on evolving the harness, which manages context, memory, tools, permissions, and runtime control, to adapt to emerging risks. It decomposes the harness into four components with clear responsibilities and introduces an evolution loop that learns from trajectory failures and selects improved harnesses through safety-utility validation.

What the Document Says

• The paper presents Safety Harness Evolution (SHE), a framework for improving the safety of large language model agents.
• SHE focuses on evolving the harness, which manages context, memory, tools, permissions, and runtime control, to adapt to emerging risks.
• SHE decomposes the harness into four artifacts: System Prompt, Rule Bank, Safety Memory, and Tool Policy.
• Each artifact has explicit safety responsibilities.
• SHE introduces an attribution-guided evolution loop that learns from trajectory failures.
• The loop converts 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.
• It achieves a 3.1x ASR reduction compared with static SafeHarness.

Key Concepts Mentioned

• Safety Harness Evolution (SHE)
• large language model agents
• harness
• trajectory failures
• attribution-guided evolution loop
• Agent-SafetyBench
• ASR

Claims Made by the Document

• SHE enhances safety through harness evolution.
• SHE achieves a 3.1x ASR reduction compared with static SafeHarness.

Evidence Used

• Experiments on Agent-SafetyBench demonstrate that SHE effectively enhances safety through harness evolution, achieving a 3.1x ASR reduction compared with static SafeHarness.

Methods / Approach

• Decomposing the harness into four artifacts
• Attribution-guided evolution loop

Limitations Stated by the Authors

• UNKNOWN (not stated in source)

Limitations Noted by the Agent

Agent interpretation — not from the source.

• The model may not be able to identify all potential limitations or weaknesses in the research.

Open Questions

• What is the exact mechanism by which SHE converts trajectory failures into structured diagnoses?
• How does SHE learn artifact-specific boundary refinements?

Notes for Future Q&A

May include interpretation. Labeled where it does.

• The paper does not provide detailed explanations on how SHE converts trajectory failures into structured diagnoses or how it learns artifact-specific boundary refinements.

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.