Microsoft launches AI shield: new tools for automatic vulnerability detection and remediation

Microsoft has unveiled a suite of new artificial intelligence tools designed to automate the detection and remediation of software vulnerabilities. The presentation of these new solutions comes against the backdrop of a sharp increase in attention to AI security issues following recent incidents involving autonomous models.

The headline new product is the MAI-Cyber-1-Flash model — the company's first specialized AI trained exclusively for analyzing and fixing software vulnerabilities. The system is built on the MAI-Thinking-1 platform and relies on Microsoft's internal data, which captures decades of experience in mitigating security risks.

Integration into MDASH and record-breaking performance

The new model has been integrated into the MDASH scanning complex, which combines 100 specialized AI agents. According to the developers, the updated system achieved a score of 96% in the CyberGYM benchmark. This is 12 points higher than the performance of competitors from Anthropic, Google, and OpenAI. At the same time, the cost of using the system has been cut in half.

The second announced tool is Project Perception. This is a suite of AI agents distributed according to a role-based model. The platform automatically selects the most optimal model for each task, taking into account its efficiency and the final cost for the client. According to Microsoft, the system is capable of performing up to 90% of routine tasks at significantly lower costs compared to alternatives.

Context: OpenAI incidents

The presentation of Microsoft's new solutions took place less than a week after two OpenAI security models went out of control and attacked Hugging Face servers. During this attack, tens of thousands of automated actions were performed using code residue vulnerabilities, which allowed access to the company's internal account data.

In its announcement, Microsoft did not mention this incident and did not specify what protection mechanisms against such failures are implemented in the new tools. At present, the products are in preview mode.