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 have proposed a framework called Safety Harness Evolution (SHE) for large language model (LLM) agents. SHE aims to improve the safety of these agents by learning evolving safe boundaries from rollout trajectories, and it breaks down the harness into four components with distinct safety responsibilities. The researchers claim that their approach effectively enhances safety through harness evolution and improves both benign utility and generalization to unseen risks.

What the Document Says

• The authors propose Safety Harness Evolution (SHE) for 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.
• The framework defines 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.
• It selects evolved harnesses through safety-utility validation.
• The experiments on Agent-SafetyBench demonstrate that SHE enhances safety and reduces ASR by 3.1x compared to static SafeHarness.
• SHE improves both benign utility and generalization to unseen risks.

Key Concepts Mentioned

• Safety Harness Evolution (SHE)
• large language model agents
• rollout trajectories
• System Prompt
• Rule Bank
• Safety Memory
• Tool Policy
• attribution-guided evolution loop
• Agent-SafetyBench
• ASR reduction

Claims Made by the Document

• SHE effectively enhances safety through harness evolution.
• SHE improves benign utility.

Evidence Used

• The experiments on Agent-SafetyBench demonstrate a 3.1x ASR reduction compared with static SafeHarness.
• The evolved harness further generalizes to unseen risks on the held-out AgentHarm benchmark.

Methods / Approach

• Learning evolving safe boundaries from rollout trajectories
• Decomposing the harness into four artifacts
• Introducing an attribution-guided evolution loop

Limitations Stated by the Authors

Not stated in source

Limitations Noted by the Agent

Agent interpretation — not from the source.

The success of SHE may depend on the quality and quantity of rollout trajectories, as well as the effectiveness of the attribution-guided evolution loop.

Open Questions

• What is the long-term impact of SHE on large language model agent safety?
• How does SHE handle complex or dynamic risks not captured in the rollout trajectories?

Notes for Future Q&A

May include interpretation. Labeled where it does.

The text focuses on the presentation of the Safety Harness Evolution framework and its performance on Agent-SafetyBench. It does not provide information about potential limitations, long-term impact, or handling complex or dynamic risks.

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.