The Pulse
Microsoft Says AI-Aided Research Helped Find 140 Windows CVEs
Microsoft says its FORGE Lab helped identify Windows vulnerabilities assigned 140 CVE numbers from May through September 2026. The company also reported 155 internally validated vulnerability reports across 23 open-source projects, while wa
AI.info Team · Written with AI ·
“Discovery creates security value only when validation and remediation can keep pace.”
Taesoo Kim, vice president of Agentic Security at Microsoft
Microsoft says its Frontier Offensive Research & Generative Exploitation Lab, or FORGE, helped find Windows vulnerabilities assigned 140 common vulnerability and exposure numbers between May and September 2026. The company says 52 were addressed in September’s security release. The results, published October 7 by Taesoo Kim, frame AI-assisted discovery as an increasingly productive source of security findings—and point to the work required to verify and fix them.
140 Windows CVEs across five months
FORGE uses MDASH, Microsoft’s multi-model agentic scanning harness, to organize vulnerability research across Windows and open-source software. Microsoft describes the Windows total as vulnerabilities FORGE helped discover that were assigned CVEs, rather than claiming that an AI system independently found and resolved all 140. Human researchers and engineering teams remain part of the process.
The accompanying monthly chart breaks the 140 vulnerabilities into servicing cohorts: 16 in May, 10 in June, 27 in July, 35 in August and 52 in September. Microsoft explicitly cautions that those monthly totals reflect announcement and servicing cohorts—not the dates the vulnerabilities were discovered or the amount of code scanned. The distinction matters: the chart tracks security releases, not a month-by-month measure of the AI system’s discovery rate.
Microsoft also describes an internal project that used deterministic algorithms, including abstract syntax trees, to reduce duplicate findings by about 45% across repeated scans of the same code. Kim argues that duplicate reports and weak evidence can consume the review capacity needed to verify stronger findings. The practical measure, he writes, is not simply how many issues a scanner flags, but how many findings arrive with evidence that lets the next team act.
Verification becomes the constraint
As automated tools generate more candidate bugs, Microsoft says, the bottleneck shifts toward establishing whether a report is real, reachable and security-relevant. A candidate may need a reproducible trigger, a working build, or proof that an attacker can reach the vulnerable path. Microsoft says project-specific proof-of-vulnerability tools and test harnesses can help produce that evidence and reduce manual triage.
The company’s proposed workflow treats verification and remediation as part of discovery: automated checks, human review, patch development and regression testing feed information back into later scans. Failed reproductions and rejected reports can also help tune the system, provided they record why a finding did not hold up. The aim is to preserve engineering accountability while automating repeatable work around the investigation.
Microsoft says the same challenges apply to open-source projects, where maintainers have different reporting rules and limited review capacity. Over three months, FORGE submitted 155 internally validated reports across 23 projects, including the Linux kernel. At the time of publication, 93 reports across 14 projects or project families had documented maintainer acknowledgement or acceptance.
Linux tests show the cost—and the caveats
Microsoft reports that its validation agents produced supporting evidence for 627 suspicious Linux-kernel findings. Among 182 confirmed crash findings, proof-of-concept generation averaged $3.61 in model cost and 21.5 minutes. In six selected cases, automated exploit generation to test local privilege-escalation potential averaged $8.56 and 25.4 minutes.
Those averages used GPT-5.5 and include attempts associated with successful cases, but exclude initial screening, candidates that failed or were filtered out, and human investigation and patch preparation. They are therefore a narrow measure of automated validation—not the full cost of finding and fixing vulnerabilities. Microsoft says one Linux report also became the first Akrites submission to result in a patch merged into the kernel.
FORGE’s figures show that AI-assisted systems can contribute to vulnerability research across Microsoft’s own software and open-source projects. They also underline the limit of counting findings alone: a vulnerability has security value only after teams can reproduce it, assess its impact and deliver a verified fix.