---
id: 5442
title: "Agent Capability Is a System Design Problem: Lessons From a 90% Improvement on CyberGym"
source_domain: depthfirst.com
source_url: "https://depthfirst.com/research/agent-capability-is-a-system-design-problem-lessons-from-a-90-improvement-on-cybergym"
published_at: "2025-11-24T00:00:00Z"
tags: ["ai-security", "benchmark", "ai-agents", "llm", "vulnerability-exploitation", "cybergym", "system-design"]
phying_url: "https://phying.com/item/5442"
lang: en
---

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

## Summary

By redesigning its agent around the model, with better context, runtime instrumentation and a modular multi-agent architecture, depthfirst lifted CyberGym exploitation success from ~28% to 53%.

## Intelligence Metadata

- **Sources:** depthfirst.com
- **Published:** 2025-11-24T00:00:00Z
- **Original:** https://depthfirst.com/research/agent-capability-is-a-system-design-problem-lessons-from-a-90-improvement-on-cybergym
- **Phying News:** https://phying.com/item/5442