- Project Perception: Microsoft is preparing an AI bug finder that could debut in July.
- Model Routing: The planned system could select among Anthropic, OpenAI, and Microsoft models for different software-security tasks.
- Security Workflow: It could find flaws and propose fixes, but security teams would still need testing and human approval.
- Product Unknowns: Microsoft has not confirmed availability, pricing, performance, customer eligibility, or final launch timing.
Microsoft is reportedly preparing Project Perception, an AI system that could find software bugs, support proposed fixes, and debut in July 2026. The unconfirmed product could combine models from Anthropic, OpenAI, and Microsoft.
Enterprise security teams could use the planned system to shorten work spanning detection, triage, patch creation, and testing. Microsoft has not confirmed the product, availability, pricing, architecture, performance, customer eligibility, or final timing.
Anthropic, OpenAI, and Microsoft models could each handle parts of the workflow. Such a design would make Microsoft partly dependent on Anthropic technology while competing with Anthropic’s Mythos AI model for enterprise security work. Routing code among providers could also make data handling and validation responsibilities harder to define.
A Three-Provider Bug-Finding System
Project Perception could assign different parts of vulnerability work to different AI models. Finding a suspicious code path is only the first step: a useful security agent must determine whether the flaw is exploitable, propose a change, and test whether that change closes the weakness without breaking the application. A model router, meaning software that chooses a model for each task, could select among OpenAI, Anthropic, and Microsoft systems instead of relying on one model throughout the workflow.
Routing work across providers is the product’s clearest proposed distinction, not evidence of better security. Microsoft may position the orchestration as a way to reduce costs, although no pricing, comparative cost data, detection benchmark, or false-positive measurement supports the positioning. Security teams would need controls for code access and retention, plus records showing which model produced each finding or proposed patch before deployment.
Human review remains necessary because software-generated fixes can alter authentication, memory handling, permissions, or other sensitive behavior. Testing must check whether a fix closes the weakness without creating regressions elsewhere, while audit records must preserve which model proposed the change and which reviewer approved it. Federal networks can require weeks or months for patching while attackers can weaponize vulnerabilities within hours, and one unpatched flaw can enable lateral movement and privilege escalation.
An AI Security Category Already in Motion
Anthropic’s Mythos would be the direct competitor. Anthropic has kept Mythos access controlled because vulnerability-finding technology can support defensive work or misuse. Project Perception’s possible reliance on Anthropic models would make Microsoft both a customer of the underlying technology and a competitor.
OpenAI expanded its Daybreak cyber-defense program on June 22. OpenAI’s broader tools-and-partnerships program spans vulnerability validation, prioritization, patch generation, and patch testing rather than stopping after a scanner flags suspicious code. Its ability to generate and test patches overlaps with some work Project Perception could target, but Daybreak is a broader program rather than one application-security agent.
Codex Security would offer a narrower comparison. OpenAI moved its predecessor, Aardvark, from an October 2025 private beta into the Codex Security name and research preview in March; limited early access is not general availability. Codex Security builds codebase knowledge, validates suspected vulnerabilities, and proposes patches for human review; by June 22, it had scanned more than 30 million commits across more than 30,000 codebases.
Google’s Big Sleep cybersecurity agent is also capable to find open-source vulnerabilities, demonstrating that finding flaws is only one part of a credible remediation workflow; security teams still need validation, tested fixes, and controlled deployment. Microsoft would need to identify supported repositories and customers, explain how code moves among providers, and publish pricing and detection results. False-positive rates, patch tests, and human-review controls would determine whether routing offers practical value rather than architectural complexity.
