Microsoft AI releases cybersecurity model MAI-Cyber-1-Flash, improving vulnerability detection performance and reducing costs by 50%

Microsoft AI has launched its first dedicated cybersecurity model, MAI-Cyber-1-Flash, and integrated it into the multi-agent vulnerability identification and remediation platform MDASH. The combination achieved a 95.95% score on the CyberGym vulnerability detection benchmark, surpassing Mythos, Gemini, and GPT series models, while reducing costs by approximately 50% compared to the existing MDASH solution.

Microsoft

Kimi K3 on PAC1 and ECOM1: results of 204 tasks and failure analysis

Moonshot AI open-sourced Kimi K3 in July. The model was tested on two BitGN benchmarks (PAC1 and ECOM1) with 204 open traces. Results show strong performance on simple retrieval and calculations but frequent failures on multi-file atomic operations, untrusted input checking, and long-table consistency.

Moonshot AI