Ridge Security Benchmark Challenges AI Model Dominance in Autonomous Red Teaming — Ridge Security Technology Inc.

Event summary

Ridge Security published the first public benchmark comparing eight AI models in autonomous penetration testing, evaluating 96 model-target test runs.
Grok 4.5 led with 77% cumulative coverage, while GPT-OSS-120B achieved the highest efficiency at 16.9 findings per million tokens.
The benchmark revealed that model performance alone doesn't determine effectiveness in autonomous security workflows.
RidgeGen™ harness demonstrated the importance of architecture in managing execution, adaptation, and validation of AI models in security tasks.
Frontier models showed alignment issues, refusing certain actions mid-workflow, highlighting the need for architectural safety controls.

The big picture

Ridge Security's benchmark underscores a shift in AI-powered cybersecurity from model selection to system architecture. As autonomous red teaming gains traction, the ability to verify and reproduce findings will become a critical differentiator. The findings challenge the assumption that the most capable general-purpose LLM will automatically excel in specialized security tasks, highlighting the need for tailored architectures like RidgeGen™. This trend reflects broader industry movements toward modular, adaptable AI systems in enterprise security.

What we're watching

Model-Agnostic Flexibility: Whether Ridge Security's model-agnostic architecture will allow organizations to adapt to evolving AI models without platform overhauls.
Cost-Efficiency Tradeoffs: How smaller and open-source models will compete with frontier models in balancing coverage, efficiency, and deployment flexibility.
Alignment Challenges: The pace at which AI models will improve in maintaining alignment with authorized security testing boundaries.