🔒AI Agents Top CyberGym: Atlas at 90.9%, MDASH Leads
Atlas and MDASH top the charts, but which one fits your security needs?
TL;DR
Wiz's Project Atlas AI agent scored a 90.9% success rate on CyberGym, while Microsoft's MDASH achieved an impressive 95.95%. Both use multi-model approaches to optimize performance and cost.
Project Atlas from Wiz and Microsoft's MDASH have topped the latest CyberGym benchmarks with success rates of 90.9% and 95.95%, respectively. These numbers matter because they highlight how different models excel in specific areas, impacting security workflows and decision-making for teams. Atlas uses Claude Opus 4.6 alongside GPT-5.5, while MDASH combines MAI-Cyber-1-Flash with GPT-5.4 to handle up to 90% of tasks efficiently.
Key Points
Project Atlas achieved a 90.9% success rate on CyberGym using Claude Opus 4.6 with GPT-5.5.
Microsoft's MDASH scored 95.95%, combining MAI-Cyber-1-Flash and GPT-5.4 for up to 90% task coverage.
Atlas routes tasks to the most efficient model, reducing costs by half compared to single-model systems.
MDASH handles routine tasks with MAI-Cyber-1-Flash, handing complex ones to GPT-5.4 for thorough validation.
Both solutions offer continuous security coverage as models evolve and improve over time.
Why It Matters
If you're managing a large codebase or dealing with frequent updates, Atlas's multi-model approach can halve your costs while maintaining high performance. For instance, MDASH’s MAI-Cyber-1-Flash handles routine tasks efficiently, reducing the need for expensive full scans.
Frequently Asked Questions
Why does this matter?
If you're managing a large codebase or dealing with frequent updates, Atlas's multi-model approach can halve your costs while maintaining high performance. For instance, MDASH’s MAI-Cyber-1-Flash handles routine tasks efficiently, reducing the need for expensive full scans.
What happened?
Wiz's Project Atlas AI agent scored a 90.9% success rate on CyberGym, while Microsoft's MDASH achieved an impressive 95.95%. Both use multi-model approaches to optimize performance and cost.
Comments
Be the first to comment
Enjoyed this article?
Get it daily. 7am. Free. Reads in 5 minutes.