Evolution Under Fire: Evolving Agent Safety Harnesses with Adaptive Injection Stressors
The paper introduces Evolution Under Fire, a staged framework that jointly evolves executable safety harnesses and adaptive injection stressors for tool-using language-model agents. A utility-constrained, Pareto-scheduled tree search evolves complete harnesses from execution evidence while the task model, benchmark, tools, and evaluator remain fixed; adaptive stressor evolution then discovers utility-preserving attacks that drive a later stage of harness evolution. On AgentDyn with GPT-4o-mini, the evolved harness cuts attack success from 50.4% to 3.0% while improving clean and attacked utility versus the initial harness. On Agent-SafetyBench with Qwen2.5-7B-Instruct, attack success falls from 82.6% to 62.0% with helpfulness and runtime validity retained, and the harness transfers without further evolution to Qwen3.5-35B-A3B, reducing attack success from 32.1% to 0.9%.
It offers a concrete, benchmarked method for automating prompt-injection defenses in tool-using agents, moving beyond hand-designed static safeguards.
Comments & Questions (0)
No comments yet
Be the first to comment!