security
May 13, 2026
By Teun
Microsoft says new AI security system found 16 Windows flaws
Microsoft said its new multi-model agentic security system, codename MDASH, helped identify 16 vulnerabilities in Windows networking and authentication components, including four critical remote code execution bugs. The company said the system also scored 88.45% on the public CyberGym benchmark and found all 21 planted bugs in a private test driver with zero false positives.
Microsoft said today that its new AI-driven security system helped researchers find 16 vulnerabilities across Windows networking and authentication components, including four critical remote code execution flaws. The company said the findings came from its multi-model agentic scanning harness, codename MDASH, which was built by Microsoft’s Autonomous Code Security team.
According to Microsoft, the bugs were found in components including the Windows kernel TCP/IP stack, IKEv2 service, dnsapi.dll, netlogon.dll, http.sys, and telnet.exe. The company said most of the issues were reachable from the network without credentials, and that the results were tied to the 5.12.2026 Patch Tuesday cohort.
⚡ New to this?
This matters because Microsoft is saying AI can now do more than spot suspicious code - it can help find real, exploitable bugs in complex system software. MDASH is an agentic system, meaning it uses multiple specialized AI agents that each do part of the security review instead of relying on one model alone.
A CVE is a public identifier for a security flaw. Remote code execution means an attacker may be able to run code on a system from a distance, while a use-after-free is a memory bug where software uses data after it has already been released.
🦞 OpenClaw angle
If you build self-hosted AI security tools, treat model choice as only one layer of the system. Split work into separate agents for scanning, debating, deduping, and proof, and add domain-specific plugins for the parts your model cannot infer, such as kernel rules or protocol state machines.
Also build in a false-positive gate before findings reach humans. Microsoft’s results suggest that validation and proof stages matter as much as detection, especially when the target is a high-value codebase where noisy output wastes a lot of reviewer time.
Microsoft described MDASH as more than a single AI model. Instead, it uses more than 100 specialized agents across multiple frontier and distilled models to scan code, debate candidate findings, and prove whether a bug is real end to end.
The company said the process is organized into stages. A prepare stage builds language-aware indices and maps the attack surface. A scan stage produces candidate findings. A validate stage has separate agents argue for and against exploitability. A dedup stage groups equivalent findings, and a prove stage tries to build triggering inputs where possible.
Microsoft said the system is designed to combine different models for different jobs. It uses a panel of models, including large models for deeper reasoning, distilled models for cheaper high-volume review, and another frontier model as a counterpoint. The company said disagreement between models can be a signal that a finding deserves more scrutiny.
The company also said the pipeline uses specialized agents for different tasks rather than trying to make one model do everything. Microsoft said the system has been built using research from past CVEs and their patches, and that its plugins can inject domain knowledge that foundation models do not have on their own, such as kernel calling conventions or lock invariants.
To test the system, Microsoft scanned StorageDrive, a private sample driver used in interviews for offensive security researchers. Microsoft said the driver contained 21 deliberately injected vulnerabilities and that MDASH found all 21 with zero false positives in that run.
Microsoft also said the system identified 16 CVEs in the current Patch Tuesday set. The list includes CVE-2026-33827 in tcpip.sys, CVE-2026-33824 in ikeext.dll, CVE-2026-41089 in netlogon.dll, and CVE-2026-41096 in dnsapi.dll.
Among the examples Microsoft highlighted was CVE-2026-33827, described as a remote unauthenticated use-after-free in tcpip.sys affecting the IPv4 receive path. The company said the issue involved incorrect lifetime management of a reference-counted Path object and could be triggered by crafted IPv4 packets carrying the SSRR option.
Microsoft also detailed CVE-2026-33824, an unauthenticated IKEv2 issue in ikeext.dll. According to the company, a crafted IKE_SA_INIT packet followed by a single fragment could trigger a double-free in the IKEEXT service, which runs as LocalSystem inside svchost.exe.
Microsoft said it measured MDASH against historical Microsoft Security Response Center cases as well. The company reported 96% recall against five years of confirmed MSRC cases in clfs.sys and 100% recall in tcpip.sys. It also said the system scored 88.45% on the public CyberGym benchmark, which contains 1,507 real-world vulnerabilities.
The company said codename MDASH is already being used by Microsoft security engineering teams and is in a limited private preview for a small set of customers.
