Google Threat Intelligence Group reported its first AI-developed zero-day in May 2026, an exploit a cybercrime group had planned to use in a mass exploitation event before Google disclosed it to the vendor.
Verizon’s 2026 Data Breach Investigations Report put vulnerability exploitation at 31% of initial access, ahead of credential abuse and phishing for the first time.
Attackers can automate parts of the chain that find a way in. Defenders may still rely on manual work to close those gaps. Among the organizations surveyed, 51% take five or more days to patch critical flaws or cannot say what their mean time to patch is.
Why headcount alone may not close the gap
Remediation has always been a volume problem running against a deadline. Both variables moved the wrong way over the past three years, and the deadline moved faster.
Verizon recorded roughly 68.7 million vulnerability instances in 2022 and 527.3 million in 2025, close to an eightfold rise. Over a shorter window, the same report found full remediation of vulnerabilities in CISA’s Known Exploited Vulnerabilities catalog dropping from 38% to 26%, while median time from detection to full remediation stretched from 32 days to 43. Curves separating at that rate cannot be pulled back together by hiring.
“Manual work is the new attack surface,” says Ryan Braunstein, Security Manager at Automox. “Every manual task, whether it’s managing tickets, tracking spreadsheets, or manually patching systems, provides extra time for an attacker to exploit something.”
The operational version of that is scheduling. Every hour a fix waits on a change window or an engineer’s availability is an hour the exposure stays open, which turns process latency into something a team can measure rather than absorb.
Automox’s 2026 State of Endpoint Management Report, produced with InformationWeek, found 43% of teams spending 10 or more hours a week on manual endpoint work, and 6% fully automated.
Mondoo’s separate 2025 remediation research lands in similar territory: 62% say there are still remediating by hand, 2% are fully automated, and just 9% are confident important vulnerabilities get fixed on time.
What gets harder as automation gets better
Every task automation absorbs can make the remaining ones weigh more. Architecture calls, incident command, and oversight of the automation itself can carry higher stakes as the volume of automated action rises, and may require judgment beyond a policy.
That is where the supply may run short. The World Economic Forum’s Global Cybersecurity Outlook 2026 ranked shortages of cybersecurity expertise second among barriers to resilience, at 57% in the public sector, with 46% of small organizations reporting insufficient expertise against 29% of large ones.
Organizational theorist Richard Claydon, whose ideas Keith Ferrazzi discusses in Forbes, names the failure mode: when output rises faster than the capacity to judge it, an organization gets busier without getting smarter. An unreviewed remediation queue is that argument in operational form.
Part of the response is happening upstream of the hiring market. Saint Leo University, an NSA and DHS designated National Center of Academic Excellence in Cyber Defense Education, will require every student from fall 2026 to earn a baseline AI certification through IBM’s SkillsBuild platform, plus at least two credentials tied to their field. It applies at every degree level, carries no cost, and is a graduation requirement rather than an elective.
“We’re pushing forward and doubling down at the same time,” said Dr. Jim Burkee, president of Saint Leo University. “We have to prepare our students to use AI and understand the tools – and in the same breath, we’re leaning harder into what’s always mattered: more human, more face-to-face teaching.”
A defender who arrives already fluent in the systems the adversary uses may need less retraining mid-career. That could change what the pipeline delivers, a slower lever than tooling but one with potential longer-term benefits.
What actually holds automation back
The barriers Automox’s 2026 report identified have little to do with whether automation works. They came back as data privacy and security implications for 46%, risk of incorrect or unauthorized changes for 44%, and lack of transparency in decision-making for 34%. Every one describes what happens when the automation gets something wrong.
The concern can scale with the deployment. A bad change on one endpoint produces a ticket. The same change pushed fleet-wide could produce an incident, which helps explain the caution.
The same research names the remedy. Practitioners asked for automatic rollback at 43%, the ability to pause or override at any time at 42%, and role-based access controls and audit logs at 42%. Those read less like preferences and more like a specification.
Urgency should still be sized to the evidence. VulnCheck’s analysis of the first half of 2026 found that of 1,061 vulnerabilities attributed to AI-assisted discovery, only 14, or 1.3%, have been confirmed exploited in the wild, roughly the same rate as vulnerabilities found by any other method. The UK’s National Cyber Security Centre assesses that fully automated, end-to-end attacks remain unlikely before 2027.
Two ledgers, one budget
What that leaves is a budgeting question rather than a technology one. Every line item strengthens one of two ledgers: the machine-speed work of finding and fixing, or the judgment work of deciding what to fix and what the automation should never touch alone.
An organization that automated remediation and staffed it with defenders trained for a pre-AI fight may have improved the arithmetic while leaving judgment underfunded. The reverse may leave capable people watching a queue they struggle to clear. The question for a board is which ledger a given investment strengthens, and whether either is running empty.
