AIAny
Icon for item

EvoSafeHarness: Evolving Model- and Domain-Specific Harnesses for Securing Agents

Synthesizes a deployment-specific safety harness (natural-language policy plus executable logic) for a frozen LLM in a target domain, using model behavior, domain specs, and adversarial review to optimize safety–utility trade-offs and transfer across agent benchmarks.

Introduction

Why this matters now Most harnesses are handcrafted and fixed, so they either over-block some models or under-protect some domains. EvoSafeHarness treats the harness as a tunable, search-able artifact: it jointly synthesizes a natural-language policy and executable enforcement logic tailored to a frozen model and target domain, guided by model rollouts, domain semantics, and adversarial review.

Key findings
  • Joint search of policy text and executable code yields stronger safety–utility frontiers than one-size-fits-all defenses: on DecodingTrust-Agent, average attack success rate (ASR) falls from 45.6% to 10.0% at a 3.3-point utility cost. This shows targeted harnesses can dramatically cut attacks with modest utility loss.
  • On AgentDojo, EvoSafeHarness attains 82.8% benign utility at 0.0% ASR, outperforming comparable defenses at the same operating point and transferring unchanged to unseen AgentDyn suites.
  • Robustness under adaptive attacks: the method keeps mean ASR below 20% under adaptive PAIR attacks with a refinement budget of 16, indicating the evolved harnesses generalize beyond benchmark-specific rules.
  • Analysis insight: domain semantics determine which safety relations and trajectory state must be monitored; model and runtime behavior determine where and how strict enforcement should be applied. In short, “what to check” is domain-driven while “how strictly to check” is model-driven.
What sets it apart
  • Co-optimization of natural-language rules and executable logic so the harness is both interpretable and enforceable.
  • Guided generation using fresh-context adversarial review to avoid brittle, benchmark-overfitted rules.
  • Designed for deployment: produces artifacts that can be validated and executed in harness pipelines and that transfer across models and related domains without retraining the base LLM.
Who should care and trade-offs

Great fit if you deploy LLM agents across heterogeneous models or domains and need a deployable, inspectable safety layer that balances utility and safety per deployment. Look elsewhere if you require zero-runtime changes to the harness (EvoSafeHarness assumes the harness can be updated) or if you cannot run the candidate-generation and validation loop (search and adversarial review add compute and engineering overhead).

Information

  • Websitearxiv.org
  • AuthorsNanxi Li, Yingzi Ma, Yulong Cao, Edward Suh, Bo Li, Dawn Song, Chaowei Xiao
  • Published date2026/09/05

Categories

More Items

Defines the Discovery Certification Protocol (DCP) to audit claims of discovery by AI research agents, converting claims into executable recovery and feedback tests. Specifies multi-gate certification, Core control requirements, and a deterministic offline verifier; validated in two controlled audits.

Evaluates whether AI agents can autonomously use Sparse Autoencoders (SAEs) to discover interpretable features and causally steer model behavior. Introduces a 20-task benchmark that has agents design contrastive probes to search a 131K+ feature dictionary in Gemma-2-9B-IT and evaluates activation rank, selectivity, and causal steering against expert references.

Translates natural-language instructions into executable programs that maintain an explicit, persistent global world state and compiles state-augmented 3D oriented bounding boxes into pixel-aligned conditioning signals for pretrained video generators. The approach decouples state evolution from rendering, enabling programmable entity control, off-screen state, and long-horizon interactive scenarios.