Microsoft has unveiled a new AI-powered cybersecurity system that combines its proprietary model with OpenAI's latest GPT-4o, claiming it scores higher than Anthropic's leading Claude model on a key benchmark while cutting costs nearly in half. The system, built directly into Microsoft Defender, is designed to tackle the growing threat of AI-driven attacks.

In benchmarks released Monday, the combined model scored 96% on CyberGym, a standard cybersecurity evaluation, compared to Anthropic's Claude 3.5 Sonnet at 84%. Microsoft says the performance gain comes from a two-tier architecture: a smaller, faster model handles about 90% of queries—detecting vulnerabilities, patching them, and verifying fixes—while the larger GPT-4o handles the remaining 10% of complex cases. According to Mustafa Suleyman, CEO of Microsoft AI, this handoff process delivers “better performance than all other models combined at 50% of the cost.”

The system uses consumption-based pricing measured in security compute units (SCUs), meaning costs scale with workload. Microsoft claims savings of nearly 50% compared to current market alternatives. Suleyman emphasized that Microsoft's decades of experience managing sensitive government and enterprise data gives it a unique advantage in training these models. “We've accrued a phenomenal amount of data combined with world-class cybersecurity expertise,” he said. The new capability enters public preview on August 3rd and will gradually roll out across Microsoft's security product lineup.

Source: CNET