Evolved agent safety harnesses cut adaptive injection success
Original title: Evolution Under Fire: Evolving Agent Safety Harnesses with Adaptive Injection Stressors
On AgentDyn with GPT-4o-mini, attack success fell from 50.4% to 3.0% while utility improved.
Evolution Under Fire pairs utility-constrained evolution of executable safety harnesses with adaptive stressor evolution, keeping the task model, benchmark, tools, and evaluator fixed. On Agent-SafetyBench with Qwen2.5-7B-Instruct, attack success fell from 82.6% to 62.0% while helpfulness and runtime validity were retained. Evolved harnesses transferred to Qwen3.5-35B-A3B without further evolution, cutting attack success from 32.1% to 0.9% and improving attacked-task utility.
- Adaptive stressors find utility-preserving attacks that guide a later evolution stage.
- The AgentDyn harness beat reproduced defenses on the security-utility trade-off.
- Clean and attacked utility rose versus the initial AgentDyn harness.
- Transfer to Qwen3.5-35B-A3B needed no further harness evolution.
From the original article
1 min read at arxiv.orgTool-using language-model agents must act on information from untrusted emails, webpages, and documents. This creates a central security challenge: malicious instructions and legitimate task directives can arrive through the same interaction channel. Existing defenses are largely hand-designed and static, making further improvements dependent on repeated manual redesign.
By Yujun Wang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.