Microsoft's MDASH Finds 16 Windows Vulnerabilities

Microsoft announced a new multi-model, agentic vulnerability discovery system codenamed MDASH, and the system contributed to identifying 16 previously unknown Windows vulnerabilities, including four critical remote code execution flaws, according to Microsoft's security blog and coverage in CSO and SiliconANGLE. The vulnerabilities were patched in Microsoft's May 12 Patch Tuesday release, and affected components named in reporting include tcpip.sys, IKEEXT, netlogon.dll, dnsapi.dll, and http.sys, per CSO. Microsoft says MDASH will open to enterprise customers in private preview in June.

Per Microsoft's blog, MDASH is a "multi-model agentic scanning harness" that orchestrates more than 100 specialized AI agents across an ensemble of frontier and distilled models to find, validate, and reproduce exploitable bugs. Microsoft published internal benchmark results showing MDASH achieved 100% detection on a private test driver called StorageDrive with 21 planted vulnerabilities and zero false positives, and reported 96% recall on clfs.sys and 100% recall on tcpip.sys across five years of confirmed Microsoft Security Response Center cases, per SiliconANGLE and Microsoft's announcement. On the public CyberGym benchmark, Microsoft reported a score of 88.45%, which several outlets, including Neowin and GeekWire, described as topping Anthropic's Mythos security-research system on the same leaderboard.

The disclosure landed alongside other vendors' AI-assisted bug hunts: SecurityWeek reported that Palo Alto Networks used Anthropic's Mythos system to find dozens of flaws in its own code, and The Register described a broader "vulnpocalypse" pattern in which Palo Alto fixed 75 flaws this month versus its usual five. CSO and SiliconANGLE also named two of the higher-severity CVEs MDASH surfaced: CVE-2026-33827, a remote use-after-free in the IPv4 stack, and CVE-2026-33824, a double-free in IKEEXT. For enterprise defenders, public disclosure of exploitable, network-reachable component bugs raises the operational importance of rapid patching and staged validation, and industry-wide adoption of agentic scanning is likely to keep increasing both the volume and technical complexity of vendor disclosures.

Enrich testing pipelines with automated validation steps that can reproduce vendor-provided exploit proofs, and track vendor-disclosed CVEs such as CVE-2026-33827 and CVE-2026-33824 through standard feeds. As AI-driven discovery increases the volume of findings across vendors, teams should evaluate automation for triage, regression testing, and staged rollouts rather than relying solely on manual review.

Watch for independent, third-party validation of Microsoft's benchmark claims, including replications on public datasets and audits of false-positive and false-negative rates, since the CyberGym score is currently a vendor-reported comparative metric. Also watch whether other vendors publish similar agentic scanning systems, whether third-party security teams adopt multi-agent pipelines of their own, and for any exploit attempts against the newly patched CVEs that would change patch-prioritization guidance.