Microsoft launches Project Perception agentic security platform and MAI-Cyber-1-Flash model, scores 95.95pct on CyberGym exploit benchmark
Microsoft $MSFT unveiled Project Perception, agentic AI (Red, Blue, Green) that find, prioritize and patch vulnerabilities. Its new MAI-Cyber-1-Flash model, paired with GPT-5.4, scored 95.95% on the CyberGym exploit benchmark vs about 83% for rivals, per Axios.
Microsoft $MSFT unveiled Project Perception, an agentic cybersecurity platform built on specialized AI agents named Red, Blue and Green that identify software vulnerabilities, determine which pose the greatest risk, and write and deploy patches, according to Axios. The company also introduced MAI-Cyber-1-Flash, the first security model it has trained in-house. The model performs roughly 95% of the work done by Microsoft's MDASH vulnerability-finding system, with more demanding tasks routed to GPT-5.4. Combined, the system scored 95.95% on the CyberGym benchmark, which measures a model's ability to generate working proof-of-concept exploits for known software vulnerabilities. Competing models, including Anthropic's Mythos and OpenAI's GPT-5.5-Cyber, scored around 83% on the same benchmark. Microsoft is making MAI-Cyber-1-Flash available through Azure AI Foundry rather than releasing it publicly, using its existing customer vetting and GPU provisioning process. The launch follows similar cybersecurity AI announcements from Google and Cisco within the same week, as security vendors race to keep pace with AI-driven attacks that can speed up vulnerability discovery and exploitation. "We're not going to let the attackers have all the productivity increase," Microsoft's corporate vice president of AI security, David Weston, told Axios. Weston said the company still has to earn the right to let its security agents operate with greater autonomy.