What Microsoft's Project Perception means for you

Microsoft is preparing a new AI tool that hunts for weaknesses in software code, and it is designed to do the job more cheaply than a similar tool from Anthropic.

According to a report by The Information, summarised by TechRepublic on 17 July 2026, Microsoft is building a platform called Project Perception. It is meant to help organisations find and repair vulnerabilities, which are flaws in software that attackers can exploit.

The system uses a "model router", a piece of software that decides which AI model should handle each task. Simpler jobs, such as listing what software a company runs, reading through log files, or navigating a codebase, go to cheaper AI models. Harder jobs, like working out how an attacker could exploit a flaw or writing the code to fix it, go to more capable models. The platform draws on models made by Microsoft, OpenAI and Anthropic.

Project Perception is aimed squarely at Anthropic's Claude Mythos 5, a tool that is good at finding vulnerabilities but costs $10 per million input tokens and $50 per million output tokens (tokens are the chunks of text AI models process), and is only available to vetted security partners. Microsoft wants to offer similar analysis at a lower price.

The new platform builds on an earlier Microsoft system called MDASH, which scored 88.45% on a benchmark called CyberGym in May 2026, ahead of Mythos's 83.1% and GPT-5.5's 81.8%.

Microsoft already sells several security products, including Defender, Sentinel, Entra and Security Copilot. Project Perception would be its first major security push since a leadership shake-up in February. According to CNBC, the company is promoting the tool's cost savings.

If you work in cybersecurity or run software that needs protecting, this points to more choice and potentially lower prices for AI-driven vulnerability scanning, an area currently dominated by expensive, restricted-access tools like Mythos. AI models are increasingly able to spot real flaws in code faster than human reviewers, and companies like Microsoft and Anthropic are competing to sell that capability to businesses that need to defend their systems.

For everyday users, there is no direct product to try yet, but the underlying trend, AI tools getting faster and cheaper at finding software bugs, could mean the apps and services you use get patched more quickly.

The platform was expected to be unveiled at a Microsoft company event later in July 2026, with a public preview to follow.