Microsoft's Project Perception uses Anthropic's own models to undercut Anthropic
Microsoft is building a cybersecurity product that uses Anthropic’s models to compete with Anthropic. The product is called Project Perception, and the irony is not subtle.
The Information broke the story on July 16, and by the weekend TechRepublic, FirstPost, and several other outlets had confirmed the details through their own sources. Here is what we know.
Project Perception is an AI vulnerability scanner. It looks across a company’s code, cloud infrastructure, and endpoints, identifies exploitable weaknesses, explains the impact, and proposes fixes. Nothing new there. The architecture is the interesting part.
A model router, not a model
Instead of sending every task to one expensive frontier model, Perception uses an orchestration layer that routes each job to the best-fit model. A cheap model handles inventory checks, log parsing, and initial triage. A frontier model only gets called when the system needs to reason through a complex exploit chain or write a remediation plan touching production code.
The models in the mix come from three sources: Microsoft’s own, OpenAI’s GPT-5.6, and Anthropic’s Claude. That last one is the detail that makes this story. Anthropic sells Mythos, a security-specialized model that has become the gold standard for AI-assisted vulnerability hunting. Mythos is also wildly expensive. TechRepublic reports its estimated API cost runs about 100 percent higher than Opus and 82 percent higher than GPT, making it one of the priciest models on the market.
Microsoft’s bet is that smart routing gets you 90 percent of Mythos’s capability at a fraction of the cost. You use Claude for the hard reasoning, a cheap Microsoft model for the grunt work, and never pay Mythos prices for tasks that do not require it.
The competitive picture
Anthropic’s Project Glasswing, the deployment program for Mythos-class security tools, already covers 150 organizations across 15 countries. US government agencies kept using Mythos even after the White House designated Anthropic a supply risk. European governments and financial institutions have been lobbying for access. The demand is real.
Microsoft sees the opening. Its July Patch Tuesday fixed a record 570 vulnerabilities with AI assistance. The company has the largest enterprise install base on the planet, and it can bundle Perception into existing security agreements rather than selling it as a standalone line item. Anthropic has the better model. Microsoft has the distribution and the cost engineering.
AI security acquisitions tripled from 10 in the first half of 2025 to 29 in the first half of 2026, according to industry tracking. CISA warned that autonomous agents are opening new attack surfaces in identity and access management. Every agent with credentials is a new vector, and defenders need tooling that moves at the same speed as AI-assisted attackers.
Who is building it
Project Perception is one of the first major initiatives under Hayete Gallot, who took over Microsoft’s security division in February. She has reportedly reorganized the business to prioritize AI-powered products and cut investment in older offerings. CEO Satya Nadella spent the spring publicly attacking OpenAI and Anthropic, accusing them of duping customers by using client data to improve their own models. Finding ways to pull enterprise customers away from those models is now explicit strategy.
Microsoft has not confirmed pricing, availability, or customer eligibility for Perception. FirstPost reports it could debut as early as this month. The company did not respond to requests for comment from any of the outlets covering the story.
The deeper pattern here is that AI security is becoming a routing problem. The frontier models are too expensive to run against an entire codebase continuously, and the cheap models are not good enough for the hard parts. Perception’s architecture treats that gap as the product. Whoever builds the best router wins, and Microsoft is betting it can build a better router than anyone selling a single model.