Your team is getting much better at finding and prioritizing vulnerabilities. But as you well know, that’s only half the battle. For all those improvements, you’re not actually seeing faster fixes.
Frontier AI tools are compressing attacker timelines to a duration that, for many, would have been unimaginable even ten years ago. Exploitation now happens in days or hours, while remediation still takes weeks or months. Why? Because the remediation model we’ve relied on for so long was never built for this pace.
Remediation is the bottleneck, not discovery
The technology available to modern security teams makes them remarkably effective at finding and prioritizing exploitable vulnerabilities. For many teams, the process is almost entirely automated and industrialized. Finding a vulnerability requires a second of automated compute.
The vulnerability remediation process, however, is complex, consequential, and built around human decisions. Right now, it is the most serious bottleneck in most teams’ vulnerability management strategy.
Every patch deployment risks breaking applications. That means organizations must implement strict change management windows, thorough manual validation, and rollback planning. Those processes stretch mean time to remediation (MTTR) from days to months.
Moreover, many organizations suffer from siloed communication and unclear responsibilities. As a result, assigning ownership and mobilizing the correct engineering or IT team can be a painfully slow process.
The 90-day patch cycle doesn’t work anymore
For years, security teams took their cue from the 90-day disclosure convention, which gave vendors roughly three months to develop, test, and release a patch.
Traditionally, that was ample time. According to zerodayclock.com, the mean time to exploit (MTTE) sat at nearly a year as recently as 2021. In 2026, that same metric fell to just over a day
Mandiant’s M-Trends 2026 report, drawn from its own incident investigations, goes further still. It puts MTTE at negative seven days. That means exploitation is, on average, happening before public disclosure, not after.
Chances are, you already know why, and you probably saw this coming. Attacker timelines have been falling steadily for years. The emergence of frontier AI models is merely the most recent, albeit the most dramatic, development in that trend. But that doesn’t change the fact that security teams must adapt to this reality. They must find a way to introduce automation into the remediation workflow. But that comes with serious caveats.
Why is AI-assisted remediation so tricky?
Vulnerability remediation is much harder to automate than discovery and prioritization.
A scanner that misidentifies a vulnerability’s severity, for example, will merely produce a wrong list. That can waste analysts’ time and point their attention to the wrong places, but it doesn’t change anything in the environment.
Mistakes in remediation, however, can have a much greater impact. Mistakenly deploying a patch, closing a port, rotating a credential, or restarting a service can break something that was actually working. That might mean a broken dependency chain, an outage, or even a compliance incident.
Part of the problem is that it’s relatively cheap and easy to validate a vulnerability against a CVE or second scanner. But security teams, especially those operating in complex environments, often can’t validate a patch won’t break a downstream dependency until they’ve already deployed it, and the dependency is already broken.
A big part of getting agentic AI remediation right is accepting that you’re never going to automate everything, and nor should you. The key is finding the parts of the workflow that will benefit most.
Where does agentic AI actually help in the remediation workflow?
Automating remediation means giving human analysts a head start on the stages that slow them down most.
Triage and enrichment
AI agents can pull context such as asset criticality, exposure, and exploit intelligence automatically and instantaneously. That’s work that would normally take an analyst several hours. An agent can even add relevant business impact and exploitability data before it reaches a human. As a result, human analysts get fully annotated findings ready for a decision.
Automated ticket creation
Rather than a security team filing a generic ticket and waiting for IT to figure out ownership, an agent can identify the responsible team based on asset ownership records. Then, it drafts the ticket with remediation steps already attached, and routes it directly to the team’s existing workflow tools.
This stage has the biggest impact on manual-handoff delay reductions, and is also the easiest to automate safely. Drafting a ticket carries none of the risk of deploying a fix.
Patch management
This is the trickiest stage to automate, and security teams should only use agents for well-understood cases with a low blast radius. For example, a missing patch on a non-critical asset with a clean rollback path.
In these cases, an agent can stage and deploy the fix without waiting for a change window slot. Crucially, higher-risk deployments, like those involving production-critical systems or unvalidated third-party patches, still route to a human for sign-off.
Validation
Once either a human or agent has deployed a fix, a separate agent can re-scan the asset and confirm the vulnerability is actually closed. That means organizations never assume a deployed patch equals a resolved risk.
Auditability and traceability are essential
Auditability and traceability are essential components of effective AI governance.
Every automated remediation action needs a record of what was changed, when, by which agent, and what the outcome was. Of course, these records are essential for compliance audits, but they also offer operational benefits.
First, when something goes wrong, teams can use the audit trail to reconstruct what happened without spending hours reverse-engineering it from scratch. Second, audit trails are the evidence base that help teams scale automation over time. To hand over more responsibility to an agent, teams need a track record demonstrating reliability.
And note: you cannot bolt auditability and traceability onto an agentic system. Teams must bake it into the models from the very start.
Remediation delays are an engineering problem
If you take one thing away from this blog, let it be this: you are never going to keep pace with modern attackers focusing solely on discovery. Remediation matters just as much, if not more.
MTTE is going to keep falling as more attackers will gain access to frontier AI models. If we want MTTR to fall along with it, without causing havoc in our environments, safe and governed agentic AI adoption, as one part of a broader push toward exposure management automation, is essential.