Network defenses need to drastically evolve to match AI-assisted attackers, but we have limited data and insight to drive this evolution. The CyberAutonomy Arena enables large scale experimentation to provide these insights into the interaction and evolution of autonomous attacker and defender systems.
Get new posts by email
2026
AI-evolved programs that score better on their benchmark can crash, slow down, or return worse results on new workloads. We built AIChilles to expose these weaknesses automatically, and found them even in a spec-driven multi-agent workflow.
Existing agentic multi-host red-teaming systems are complicated and difficult to develop. To solve this, we created CyBlocks, a visual coding tool for developing AMR systems quickly and legibly.
We ran GLM 5.2, Kimi K3, and Qwen 3.8 Max on our cyber range benchmark MHBench. Qwen 3.8 Max leads in performance across all harnesses. Qwen 3.8 equipped with the Incalmo harness had the highest overall performance.