Microsoft preps Project Perception to rival Anthropic's Mythos

What happened: Microsoft is building Project Perception, an AI cybersecurity platform that uses multiple AI models to detect and repair software vulnerabilities, The Information reported, as summarised by TechRepublic on 17 July.

The details: A model router assigns each task to the best-suited model. Cheaper models handle inventory, log parsing, codebase navigation and triage. More capable models take on exploit-path reasoning and writing fixes. The platform draws on models from Microsoft, OpenAI and Anthropic.

The numbers: Microsoft's earlier MDASH system scored 88.45% on the CyberGym benchmark in May 2026, ahead of Anthropic's Claude Mythos 5 (83.1%) and GPT-5.5 (81.8%). Mythos is listed at $10 per million input tokens and $50 per million output tokens.

The rivals: Project Perception is positioned directly against Mythos, which Microsoft says is strong at vulnerability-hunting but expensive and limited to vetted security partners. Microsoft wants to match that analysis at lower cost.

Why it matters: AI models are now finding real software flaws faster than humans. Whoever sells that capability to defenders cheaply and broadly gains an edge in enterprise security.

Who's affected: Microsoft already sells Defender, Sentinel, Entra and Security Copilot to enterprise customers. Project Perception would be its first major security push since a leadership shake-up in February.

What's next: The platform was expected to be unveiled at a company event later in July, with a public preview to follow, according to CNBC.