AI Is Finding Vulnerabilities Faster Than Teams Can Fix Them

For years, vulnerability management has followed the same pattern: scan the environment, identify weaknesses, assign severity, work through the queue, and repeat.

AI is changing the scale of that problem. Vulnerabilities can now be discovered and analyzed faster, but the work that follows still depends on people, testing, maintenance windows, application dependencies, and the operational risk of making changes in production.

Microsoft recently warned that frontier AI models are increasing the pace and volume of vulnerability discovery. September 2026 alone saw close to 1,000 vulnerabilities released through Microsoft’s security update process for on-premises software. Microsoft is also using AI internally to scan its own codebase and expects this type of scanning to become part of standard security assurance over time.

For security leaders, the challenge is increasingly less about finding vulnerabilities and more about deciding which findings deserve attention first, which systems require immediate action, and how quickly remediation can happen without creating a different operational problem.

More findings put pressure on an already stretched process

Most vulnerability programs were built for a relatively predictable flow of findings and patch cycles. Infrastructure teams still need time to test updates, coordinate dependencies, schedule maintenance windows, and account for systems that cannot be taken down casually.

If vulnerability discovery accelerates while remediation capacity stays roughly the same, the likely result is a larger backlog. This is especially relevant for organizations with substantial on-premises infrastructure, because Microsoft is telling CISOs to expect a higher volume of vulnerabilities in on-premises software.

That does not mean every critical CVE should trigger an emergency change. It does mean the organization should already know what qualifies for expedited treatment, who can approve it, what testing is required, how rollback will work, and what controls are available if the patch has to wait.

Severity alone doesn’t tell you what to fix first

A critical CVSS score is useful information, but it does not tell you enough about actual exposure.

The same vulnerability can represent very different levels of risk depending on where the system sits, whether it is internet-facing, what identities can access it, what privileges are available from it, whether exploitation has been observed, and whether an attacker can use it as part of a broader path through the environment.

Vulnerability management increasingly needs to be connected with identity, asset criticality, configuration, external exposure, threat intelligence, and attack-path information.

Microsoft Security Exposure Management is one example. Its value is not simply finding another vulnerability, but adding enough environmental context to understand whether a weakness sits on a realistic route to a privileged identity, sensitive data, or a critical workload.

Many organizations still struggle here. Security teams often pass infrastructure teams a list of CVEs and severity scores without enough context about the affected service or likely consequences of exploitation. That makes prioritization harder and can leave teams spending time on findings that are technically severe but less likely to be used in a real attack.

Patch timing may need to change

Microsoft also recommends reconsidering patch timing for critical systems.

Many organizations intentionally patch domain controllers, edge devices, and other sensitive infrastructure during evenings, weekends, or established maintenance windows because a failed update can create significant disruption.

The calculation changes if AI shortens the time between patch release and working exploitation.

Microsoft suggests considering fixes to critical systems within 24 hours in some circumstances rather than automatically waiting for the next maintenance window. That should not become a universal 24-hour patch requirement. It is a reason to review whether existing change processes can respond quickly enough when a vulnerability presents an immediate risk.

That decision is easier when critical assets are already classified, emergency change procedures are tested, and ownership between security and infrastructure teams is clear.

Supporting reference: CISO perspectives on managing vulnerability risks in the age of AI ↗

Defense in depth still buys time

No organization is going to patch every vulnerability immediately. Legacy applications, vendor dependencies, unsupported systems, change freezes, and operational requirements will continue to get in the way.

The controls around a vulnerable system therefore remain important. Phishing-resistant MFA for privileged users, Conditional Access, endpoint detection and response, segmentation, attack surface reduction, least privilege, and effective monitoring can all reduce the likelihood that a single vulnerability turns into a broader compromise.

Microsoft also points security leaders toward Baseline Security Mode as a way to apply and monitor secure configuration across Microsoft environments. Stronger baselines reduce the number of weaknesses an attacker can combine after initial access.

The difficult part is applying those controls in production. Dependencies need testing, exceptions need owners, and temporary exceptions need to be revisited rather than quietly becoming permanent. Anyone who has worked through a hardening exercise in a live environment knows that turning on the control is usually the easy part. Understanding what breaks and why takes more work.

Microsoft is already building for this problem

Project Perception gives us an indication of where Microsoft is taking this.

It is a coordinated multi-agent security system that brings together red team agents that look for weaknesses, blue team agents that investigate threats, and green team agents that harden the environment. Its first announced scenario is software vulnerability management, where Microsoft is bringing its MAI-Cyber-1-Flash model into MDASH, the company’s multi-model vulnerability scanning system.

That connection is useful because MDASH is also referenced in Microsoft’s latest CISO guidance. Microsoft describes it as a harness around AI models that controls how they access code, validate findings, and feed results into triage and remediation. Microsoft has expanded its use internally and has started making the capability available to customers.

If AI starts producing credible findings faster than teams can investigate them, improving the scanner alone will not solve the problem.

Coordinated agents can help validate findings, add context, support prioritization, and move remediation forward while keeping people involved in consequential actions. The goal should be fewer manual handoffs and less time spent validating findings, not another queue for the SOC to work through.

Supporting reference: Project Perception ↗

AI should reduce repetitive analysis

AI can already help with parts of vulnerability management without taking over the whole process.

It can correlate findings with affected assets, summarize exploitation information, draft remediation guidance, and support validation after a change. Used carefully, that reduces repetitive analysis and gives teams more time for systems that require human judgment.

The workflow around the model still matters. Findings need validation, remediation still needs change control, exceptions still need owners, and someone still needs to decide whether the operational risk of applying a fix is lower than the security risk of waiting.

AI can help teams get to those decisions faster, but someone still owns the risk.

The existing process needs to keep up

For most organizations, the starting point is not a wholesale redesign of vulnerability management.

Routine patching and expedited remediation should be treated differently. Critical assets should be identified before an emergency occurs, and vulnerability data should be connected to identity, asset importance, internet exposure, and attack paths. Exceptions should have both an owner and an expiration point.

Microsoft’s work with MDASH and Project Perception shows that the defensive side is already moving toward more automated validation, prioritization, and remediation support.

For many security teams, the priority is shortening the time between a credible finding, a risk-informed decision, and safe remediation. That means knowing what is exposed, who owns the decision, and how quickly the organization can safely act.

Vulnerability data is rarely the only constraint. Teams need a process that identifies what requires attention now and can move safely when the risk is high enough.