Project Ire: Microsoft’s AI System That Hunts Malware Without Humans
Microsoft has unveiled an autonomous artificial intelligence system capable of analyzing and classifying software without any human intervention. The prototype, named Project Ire, is designed to detect malicious code at scale.
According to Microsoft Research’s official blog, the system performs what is referred to as the “gold standard” of analysis—it fully reverses an executable file without any prior knowledge of its purpose or origin. To accomplish this, it employs a comprehensive arsenal of tools: decompilers, API interfaces, memory sandboxes including Project Freta, as well as both open-source and proprietary solutions.
The classification process unfolds in several stages: automated reverse engineering, reconstruction of the control flow graph using tools such as Ghidra and angr, semantic analysis of code behavior, and invocation of specialized validation tools. All system logic is meticulously documented in a reasoning log—a “chain of evidence” that allows for the verification of the final verdict.
In one test, Project Ire was evaluated on a dataset of Windows drivers that included both malicious samples from the Living off the Land Drivers collection and legitimate ones from Windows Update. In this scenario, the system successfully classified 90% of the files, achieving a precision score of 0.98, a recall of 0.83, and a false positive rate of just 2%.
A more challenging test involved nearly 4,000 files that had eluded classification by automated tools and were intended for manual analysis by experts. Project Ire operated in a fully autonomous mode, processing files created after the training of the language models—at a time when no other Microsoft automation tool could classify them.
In this case, the precision was 0.89—meaning nearly 9 out of 10 files flagged as malicious were indeed malicious. The recall, however, was 0.26, indicating that about a quarter of all malicious samples were successfully detected. Nonetheless, the false positive rate remained low at just 4%.
Though the overall performance in this test was moderate, the combination of high precision and minimal error rate underscores the system’s tangible potential for future deployment.
Moving forward, the prototype will be integrated into the Microsoft Defender platform under the name Binary Analyzer, where it will serve as an internal tool for automated binary file assessment. The project’s goal is to increase both the speed and accuracy of classifying novel software, including previously unseen programs.
Additionally, Microsoft has released the results of its vulnerability reward program. According to the annual report by the Microsoft Security Response Center, the company paid out $17 million to 344 researchers from 59 countries over the span of a year. Between July 2024 and June 2025, a total of 1,469 valid reports were submitted, with the highest single payout reaching $200,000.
