Agent Capability Is a System Design Problem: Lessons From a 90% Improvement on CyberGym

depthfirst · Achievement · November 24, 2025

The company announced a significant improvement in the ability of AI agents to exploit real-world software vulnerabilities. By optimizing environment design and multi-agent architecture, the system reached a 53% success rate on the CyberGym benchmark.

Read the announcement

AntSpace’s summary of depthfirst’s own announcement: read the original.

More updates from depthfirst