Microsoft is developing an internal security platform codenamed Project Perception that aims to use multiple AI models to find, verify, and fix software vulnerabilities at scale. The effort, first reported by The Information, could position the company as a direct competitor to Anthropic’s closely held Claude Mythos program, but with a multi-model architecture designed to cut costs and broaden access.

Project Perception has not been officially announced. No release date, pricing, or integration details exist. Yet the reporting, combined with Microsoft’s publicly demonstrated research, paints a clear picture of where enterprise vulnerability management is heading—and what Windows security teams should expect.

What We Know About Project Perception

The core idea is straightforward: instead of relying on a single frontier model to scan codebases, Microsoft is building an orchestration layer that routes vulnerability-discovery tasks to different AI models depending on complexity and cost. According to TechRepublic and The Information, the system will combine models from Microsoft, OpenAI, and Anthropic, assigning simpler jobs like code indexing or log parsing to cheaper models while reserving more capable models for exploit reasoning or remediation guidance.

This design stems from Microsoft’s Multi-Model Agentic Scanning Harness, or MDASH, which the company detailed in a May 2026 security blog post. MDASH uses over 100 coordinated AI agents to autonomously discover, reproduce, validate, and prioritize software bugs. It does not simply generate plausible-sounding vulnerability reports; it attempts to prove the vulnerable path exists, determine real-world exploitability, and produce a defensible finding.

Microsoft reported that MDASH found 16 previously unknown vulnerabilities in Windows components, including four critical remote-code-execution flaws. In testing against five years of confirmed Microsoft Security Response Center cases, the system achieved 96% recall on clfs.sys and 100% on tcpip.sys—without false positives. On the public CyberGym benchmark, MDASH scored 88.45% across 1,507 vulnerability-reproduction tasks drawn from 188 OSS-Fuzz projects, roughly five percentage points ahead of the next entry.

These are vendor-reported research results, not a guarantee of future product performance. Still, they demonstrate that a multi-model, agentic approach can produce actionable security findings with a high degree of confidence.

Project Perception reportedly targets the same enterprise market as Anthropic’s Claude Mythos, a frontier model tested through the invitation-only Project Glasswing. Mythos has demonstrated extraordinary bug-finding capabilities—more than 10,000 high- and critical-severity vulnerabilities discovered among roughly 50 partners—but it remains tightly controlled due to offensive potential and high operational cost. Microsoft’s multi-model strategy offers a path to similar results without the same price tag or access restrictions.

What Project Perception Means for Windows Security Teams (and Developers)

There is no new dashboard to deploy today. Project Perception remains a report, not a product. But the direction is significant.

For Windows administrators and security analysts, the most meaningful shift is conceptual: vulnerability management is moving from periodic scanning followed by manual triage to continuous, AI-driven detection with built-in verification. A system that not only flags a suspicious code path but also reproduces the issue, assigns ownership, and suggests a tested fix could transform overburdened security teams.

Consider internal business applications, PowerShell automation, custom drivers, or the ad-hoc integrations that connect Active Directory to cloud services. Conventional scanners often miss logic flaws or multi-step exploit chains in such code. An agentic system like MDASH can reason across components and data flows, potentially uncovering the kind of vulnerabilities that are most dangerous in a Windows enterprise.

But the promise of fewer false positives comes with new concerns. Multi-model routing means source code and findings could be processed by different AI services, each with its own data-handling policies. Security leaders will need clear answers: Which model sees proprietary code? Is processing done in a customer’s tenant or in a shared environment? Are prompts and intermediate results retained or used for training? Can administrators force on-premises processing for sensitive repositories? And if an underlying model changes, can an earlier finding be reproduced exactly?

Microsoft’s existing security portfolio—Defender XDR, Microsoft Sentinel, Azure, GitHub Advanced Security, Security Copilot—gives it a natural advantage in answering some of these questions. Deep integration with Windows telemetry, identity, and deployment pipelines could make Project Perception far more than a standalone bug finder. It could become a closed-loop remediation engine that validates patches in a sandbox before rollout. But that level of automation requires ironclad controls to prevent automated remediation from becoming automated disruption.

For developers, the tool could eventually appear in familiar workflows. GitHub Advanced Security already performs static analysis and dependency scanning. A future integration might flag a potential vulnerability, provide a proof-of-concept exploit, and even generate a tested pull request with a fix. That would shift security left in a profoundly tangible way—if the underlying AI’s confidence is high enough.

The Context: Why Microsoft Is Building This Now

The urgency is driven by three forces. First, attackers are using AI to find and weaponize vulnerabilities faster than ever. The window from disclosure to exploitation keeps shrinking, and manual analysis cannot keep up. Second, Anthropic’s Glasswing program has shown that frontier models can discover vulnerabilities at industrial scale, but that success creates a new problem: the flood of findings overwhelms the patch pipeline. In its May 2026 progress report, Anthropic noted that of 1,752 high- or critical-severity findings independently assessed, 90.6% were valid and 62.4% remained high or critical after review. Yet maintainers managed to patch only one such issue every two weeks on average.

That bottleneck is the real crisis. An AI that finds ten times more bugs is only a defensive breakthrough if the organization can fix them. Microsoft’s MDASH research explicitly addresses this: the system does not stop at discovery; it prioritizes issues that are reachable and exploitable, producing evidence and proposed fixes. That verification-first approach could help teams focus on what matters most, avoiding the triage paralysis that Anthropic’s reporting describes.

Third, geopolitical pressure is mounting. Chinese AI developers are narrowing the gap with U.S. models, and governments are tightening controls on frontier AI systems. A vulnerability-management platform that combines multiple models—including Anthropic’s—diversifies supply risk and could become a strategic asset for Western enterprises and governments.

What You Should Do Now

Project Perception isn’t here yet, but the assumptions it makes about AI-driven security will likely shape the tools you’ll use within two years. Steps to take today:

  • Audit your vulnerability triage process. If your team spends 80% of its time dismissing false positives, an AI system that pre-validates findings could be transformative—but only if it doesn’t introduce its own noise. Understand your current signal-to-noise ratio and where automation would help most.
  • Map your sensitive code surfaces. What parts of your codebase, scripts, and configuration could you safely expose to an external AI analysis? What must stay on-premises? Start a data-classification project so you can move quickly when tools require clear boundaries.
  • Watch for Microsoft announcements. The most natural entry points are GitHub Advanced Security and Security Copilot. Expect a private preview or pilot within the next 12 months, likely tied to Azure Government or large enterprise agreements.
  • Plan for faster patch cycles. Even without AI tools, the industry is moving toward continuous delivery of security fixes. If your change-management process requires weeks of manual approval, begin streamlining now. When AI-generated fixes arrive, the bottleneck will be your deployment pipeline, not the code.
  • Engage with the MDASH research paper. Microsoft’s blog post is publicly available and explains the multi-model architecture in detail. Understanding its design principles will help you evaluate future products.

Outlook

Project Perception represents a strategic bet that the future of vulnerability management is not a smarter scanner but a smarter orchestration of many specialized agents. By building a control plane above the models, Microsoft aims to own the verification layer—the place where raw AI output becomes a trusted, actionable security finding.

The competition with Anthropic is less about who finds more bugs and more about who builds the most defensible and manageable remediation pipeline. Anthropic’s Glasswing data already shows that volume alone can break the process. Microsoft’s MDASH results suggest a path to precision.

For Windows shops, the likely outcome is a security platform that grows more autonomous and more integrated. Vulnerability hunting may soon feel less like a manual chore and more like a collaborative triage between human analysts and an orchestra of AI agents—each playing its part, but only after the conductor verifies the tune.