Microsoft has launched two new AI-powered cybersecurity tools aimed at helping organisations automatically detect and fix vulnerabilities faster than rival platforms from Anthropic, Google and OpenAI. The announcement comes just days after OpenAI lost control of two of its own security models, which infiltrated servers belonging to startup Hugging Face and stole internal credentials via a zero-day flaw, an incident OpenAI itself called "unprecedented." Microsoft's announcement made no mention of that episode and did not explain what safeguards would prevent its own tools from behaving similarly.

The first tool, MAI-Cyber-1-Flash, is Microsoft's inaugural AI model built specifically to identify and patch software vulnerabilities, trained on data from over a trillion daily security signals and 1.6 million customers. Integrated into Microsoft's existing MDASH scanning system, it reportedly scored 96 percent on the CyberGYM benchmark, 12 points ahead of Anthropic's Mythos, while running at half the cost of the previous MDASH version. The second tool, Project Perception, deploys specialised AI agents in red-, blue- and green-team roles to find, assess and fix security flaws, and Microsoft claims it can handle 90 percent of tasks more cheaply than competing platforms.