Evolution Under Fire: Evolving Agent Safety Harnesses with Adaptive Injection Stressors
An evolved safety harness cut attack success from 50.4% to 3.0% on AgentDyn with GPT-4o-mini while improving task utility.
The framework evolves executable safety harnesses from execution evidence, then develops adaptive attacks to guide another round of improvement. It keeps the task model, benchmark, tools and evaluator fixed. Results varied: attack success on Agent-SafetyBench with Qwen2.5-7B-Instruct fell from 82.6% to 62.0%, with helpfulness and runtime validity retained. Harnesses transferred to Qwen3.5-35B-A3B without further evolution, cutting attack success from 32.1% to 0.9% and improving attacked-task utility.
ArXiv · AI/CL/LG's note
The framework evolves executable safety harnesses from execution evidence, then develops adaptive attacks to guide another round of improvement. It keeps the task model, benchmark, tools and evaluator fixed. Results varied: attack success on Agent-SafetyBench with Qwen2.5-7B-Instruct fell from 82.6% to 62.0%, with helpfulness and runtime validity retained. Harnesses transferred to Qwen3.5-35B-A3B without further evolution, cutting attack success from 32.1% to 0.9% and improving attacked-task utility.
ArXiv · AI/CL/LG's note
score 5