Microsoft, Palo Alto Networks Find Many Vulnerabilities by Using AI on Their Own Code

"Microsoft said on Tuesday that more than a dozen of the 137 vulnerabilities fixed with its latest Patch Tuesday updates were found by a new AI system called MDASH (multi-model agentic scanning harness) built by its Autonomous Code Security team."

"MDASH is designed to run a structured pipeline that moves findings through several distinct stages: preparation, scanning, validation, deduplication, and proof construction. Different agents handle different roles: some identify candidate vulnerabilities, others argue for or against their exploitability, and a final stage attempts to construct inputs that actually trigger the bug."

"This multi-stage debate architecture means that a finding must withstand scrutiny before it reaches a human engineer. According to Microsoft, MDASH was used to discover 16 of the vulnerabilities fixed with the latest Patch Tuesday updates. Four of them were rated critical, including unauthenticated remote code execution flaws in components such as the Windows kernel TCP/IP stack and the IKEv2 serv"

"Palo Alto Networks revealed on Wednesday that it has used Claude Mythos and other frontier AI models to conduct a deep scan of its product portfolio, which resulted in the discovery of dozens of vulnerabilities."

Microsoft reported that its MDASH system, which orchestrates more than 100 specialized AI agents across multiple AI models, helped discover 16 vulnerabilities among those fixed in its latest Patch Tuesday updates. MDASH runs a structured pipeline with preparation, scanning, validation, deduplication, and proof construction stages. Different agents identify candidate vulnerabilities, assess exploitability, and attempt to construct inputs that trigger the bugs. Microsoft said the multi-stage debate architecture requires findings to withstand scrutiny before reaching human engineers. Palo Alto Networks reported using Claude Mythos and other frontier AI models to deep-scan its product portfolio, resulting in dozens of vulnerabilities. The reports reflect ongoing debate about how AI will change vulnerability discovery performance and workflows.

Read at SecurityWeek