Skip to content
InfoResearchPreprintLLM-specific

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

Published
Record updated
View JSON

Summary

Evolution Under Fire is a staged framework that co-evolves executable safety harnesses for tool-using LLM agents together with adaptive injection stressors, keeping the task model, benchmark, tools and evaluator fixed. On AgentDyn with GPT-4o-mini, an evolved harness reduces attack success from 50.4% to 3.0% while improving utility. On Agent-SafetyBench with Qwen2.5-7B-Instruct, attack success falls from 82.6% to 62.0%, and harnesses transfer to Qwen3.5-35B-A3B, cutting it from 32.1% to 0.9%.