AgentSafeEvo: Adversarial co-evolution framework for agent safety
Start here
arXiv released AgentSafeEvo, an adversarial co-evolutionary reinforcement learning framework for LLM agents: the attacker learns to construct adversarial tasks that expose weaknesses in the defender's tool use, while the defender learns to complete tasks safely under these attacks. The framework introduces learnability-guided task weighting, which adjusts each task's contribution to the updates of both attacker and defender based on the variance of the defender's reward within the task. Experiments on agent safety and general capability benchmarks show that the method improves adversarial robustness while maintaining general agent capabilities.
Generated by AI from reports · Updated 29 days ago
Attention and developments
Attention, hourly
Now 23, 0% in 24 hours·Peak 23 (Nov 6 03:00)
Related research
This event's paper
- AgentSafeEvo: Adversarial Co-Evolution for LLM Agent Safety
arXiv · 2026-10-14 · Deep reading
Progress timeline
- XiaohongshuXiaohongshu · AI Security Heart· DiscussionAgentSafeEvo: Co-evolution of LLM agent security and attacks(opens the original post in a new tab)
- arXiv: Adversarial Attacks, Backdoors & PoisoningAgentSafeEvo: An Adversarial Co-Evolutionary Framework for LLM Agent Safety
AgentSafeEvo is an adversarial co-evolutionary reinforcement learning framework for LLM agents, where the attacker learns to construct adversarial tasks that expose weaknesses in the defender's tool use, and the defender learns to complete tasks safely under these attacks. The framework introduces learnability-guided task weighting, adjusting each task's contribution to updates of both attacker and defender based on the variance of defender rewards within the task. Experiments on agent safety and general capability benchmarks show that the method improves adversarial robustness while maintaining general agent capabilities.
Related discussion
- Older than 72 hours · Not included in the 72-hour figures
- XiaohongshuXiaohongshu · AI Security HeartAgentSafeEvo: Co-evolution of LLM agent security and attacks(opens the original post in a new tab)