Microsoft is preparing a new AI security tool internally named Project Perception, according to a report by The Information. The system is designed to automatically detect and fix vulnerabilities in enterprise software, combining models from Microsoft, OpenAI, and Anthropic. It thus competes directly with Anthropic’s own security product, Mythos.
Report Describes Cost Advantage Over Mythos
As The Information reports, citing internal sources, Project Perception is expected to launch later this month. The tool selectively chooses a suitable AI model for each analysis task instead of consistently relying on expensive top models, and directs simple checks to cheaper models. This control is expected to significantly reduce costs compared to Mythos: Mythos queries reportedly cost up to 100 percent more than comparable queries with Opus and up to 82 percent more than with GPT models – unverified independently.
Project Perception is also expected to be more widely accessible than Mythos, which is currently only available to around 50 partner organizations from the so-called Glasswing program. Microsoft is pursuing an approach that other providers also apply with cheaper models from China: combining multiple systems instead of relying on a single expensive top model. There is currently no official confirmation from Microsoft, specific pricing, or a fixed launch date – all information comes from the report by The Information.
Predecessor MDASH Already Provides Verified Security Results
The technical foundation for Project Perception is provided by the internal system MDASH, which Microsoft documented again in a blog post in June. As Beckmann.ai reported back in July, Microsoft has been expanding the use of MDASH to the entire Windows code since July 9, 2026. According to Microsoft, MDASH achieved a hit rate of 96.5 percent on the CyberGym benchmark with 1,507 real vulnerability tasks. During the June Patch Day, the system identified ten CVE entries, including three critical vulnerabilities with privilege escalation in Hyper-V, as well as additional gaps in the Windows kernel, DNS and DHCP clients, and the Remote Desktop client. The Windows security team states that MDASH enables vulnerability detection on the scale of Windows with significantly greater depth than previous methods.
On the other hand, Anthropic markets Mythos as a frontier model specifically for security teams and engages companies through the Glasswing program. By May 22, 2026, the approximately 50 participating organizations identified more than 10,000 critical and high-risk vulnerabilities collectively, according to program data. Project Perception would thus package the MDASH technology for the first time as a standalone product accessible to paying customers, directly competing with this established result. For companies with their own software development, the competition in AI-supported vulnerability detection shifts from pure detection accuracy to cost per scan and accessibility. Larger IT departments may soon choose between several specialized providers instead of predominantly relying on a single system as before.
It remains to be seen whether Microsoft will meet the rumored launch date in July and how the company will publicly substantiate the results of Project Perception once the tool is actually available. The known cost advantages so far come solely from the report by The Information; an independent verification of the hit rate outside of Microsoft’s own systems is still pending.
