Microsoft report says AI accelerates familiar cyberattack methods

Microsoft’s 2026 Digital Defense Report, released October 1, says AI is accelerating cyberattack workflows while most complex intrusions still involve meaningful human direction. Its findings make a case for faster remediation and stronger controls over the identities and systems that attackers already exploit.

Microsoft puts the median time from vulnerability discovery in the wild to weaponization below 24 hours. Remediation of critical external vulnerabilities can take enterprises 30 to 60 days. These are broad figures describing different stages of security work, rather than a measured outcome for every organization.

A controlled test of attack automation

The full report generally covers July 2025 through June 2026, with exceptions identified in individual sections. Its findings describe Microsoft’s observations and evaluations.

One evaluation connected 32 attack stages in an emulated enterprise. The models completed the chain without human support, but the mock environment had no defenders. That result demonstrates capability under the test conditions. It does not establish how often an automated attack would succeed against an actively defended organization.

Familiar entry points remain important

In Microsoft Defender Experts customer notifications, user execution accounted for 30% of observed initial access and valid accounts for 20%. These figures concern the company’s observed activity, rather than every intrusion worldwide.

Among detections tied to the five leading software vulnerabilities analyzed, 58% involved a Netlogon issue disclosed in 2020. The report explicitly cautions that detection events do not prove successful exploitation. The finding concerns a narrow set of detections, not 58% of all cyberattacks.

What this means for enterprise agents

In accompanying guidance, Microsoft deputy CISO Terrell Cox argues that security depends on the whole system around an AI model. An agent’s reach includes its data, applications, tools, permissions and supporting infrastructure. The report examines agent identity, authentication between agents, attribution and the ability to withdraw access.

For organizations deploying agents, that makes access design a practical starting point. Teams need to understand what each agent can reach and how that access can be revoked. Cox identifies least privilege, monitoring, testing and secure software development as established disciplines that also apply to AI systems.

Human expertise remains part of that work. Cox says automation can handle established techniques and connect known information, while identifying undocumented attack paths still benefits from experienced operators. Faster automated work therefore leaves a continuing need for people who can investigate context and judge unfamiliar weaknesses.

ByteForward’s coverage of an auditable AI security assessment proposal examines a related question about making the evidence behind security reviews easier to inspect.

Illustrative photograph Almost done by alq666, shared under Creative Commons Attribution ShareAlike 2.0. This resized 2007 photograph illustrates computing infrastructure.