AI will not replace cybersecurity teams. It is taking over alert triage, evidence collection, control testing and first-draft reports, and that changes the job. People keep the judgment about what a finding means, the decision to accept a risk and the accountability when AI is wrong, and the 2026 surveys show them spending more of their time there.
I take that position because of one number. In May 2026, ISC2 surveyed 856 cybersecurity professionals who use AI in their roles, and 65% said they now spend more time deciding when to trust or act on AI-generated recommendations. AI took on more of the work, and people spent more of their time judging what it produced.
What an AI-augmented security team is
An AI-augmented security team is one in which AI collects evidence, tests controls, triages alerts and drafts reports, while named people decide what the findings mean, which risks the organization accepts and who answers for the outcome. That is the model the research below describes. It is also the model I would build toward, whatever tools a team buys.
What the 2026 research says
Seven studies answer the question from different directions: practitioner surveys, payroll records, a benchmark of AI agents and a study of live security operations. I have labeled vendor research as vendor research.
| Source | Who or what was studied | What it found |
|---|---|---|
| ISC2, July 2026 | 856 cybersecurity professionals who use AI, surveyed in May 2026 | 65% spend more time deciding when to trust AI recommendations, 63% spend more time validating AI outputs, and 89% have seen AI recommendations lead to incorrect outcomes. |
| SANS Institute and GIAC, March 2026 | 947 respondents in six regions, 72% of them in security leadership | 74% say AI is changing team size and role structures, 49% report less manual analysis time, and 16% report an actual headcount reduction. |
| ISACA, September 2026 | More than 1,800 cybersecurity professionals | 8% run AI-specific incident response exercises regularly, and 64% have never run one. |
| World Economic Forum, January 2026 | 804 qualified participants in 92 countries | 64% have a process to assess the security of AI tools before deploying them, up from 37% a year earlier. |
| Stanford Digital Economy Lab, revised August 2026 | ADP payroll records for millions of US workers through June 2026 | Employment of workers aged 22 to 25 in AI-exposed occupations is 19% below where it would be had it kept pace with less-exposed peers, and the authors find no evidence of economy-wide displacement. |
| Carnegie Mellon University, revised September 2025 | 175 tasks in a simulated software company | The strongest AI agent completed 30% of the tasks on its own. |
| Microsoft, November 2024, vendor research | Organizations whose security teams adopted Security Copilot, compared with a control group | Adoption was associated with a 30.13% reduction in mean time to resolve security incidents. |
Read together, the research says the work has moved from producing to checking. SANS reads its own data the same way: the main effect of AI is efficiency, and only 16% of organizations report cutting headcount. A team that uses AI well spends less time assembling evidence and more time challenging it, and the ISC2 figure of 89% is the reason the challenge cannot be skipped.
Capability is also a different thing from replacement. MIT’s Project Iceberg, built with Oak Ridge National Laboratory, estimates that AI can technically perform tasks worth 11.7% of US wage value, about $1.2 trillion, and its authors state that the index measures technical exposure rather than displacement. Carnegie Mellon’s benchmark shows the gap from the other side. An agent that completes 30% of the tasks in a simulated workplace leaves 70% for people, and someone still has to check the 30% it finished.
Which work moves to AI, and which stays with people
| Work | Who does it in an AI-augmented team | Why |
|---|---|---|
| Alert triage and log analysis | AI does the first pass, and a person reviews what it escalates. | ISC2 lists both among the tasks increasingly performed or accelerated by AI tools. |
| Evidence collection and control testing | AI tests each control on a set frequency and keeps the evidence. | A control answer written once describes the day it was written. |
| First-draft reports and board narratives | AI drafts, and a named person signs. | The person who signs is the one a board or regulator will question. |
| Deciding what a finding means | People | The context usually sits outside any system the model can read, such as a product launch or a supplier’s contract terms. |
| Accepting a risk | A named person | NIST AI RMF Manage 1.3 lists accepting as one risk response, and Govern 2.3 puts decisions about AI risks with executive leadership. |
| Deciding whether an incident is material | People | Under the SEC rules adopted in July 2023, a material incident is disclosed on Form 8-K Item 1.05 within 4 business days of that determination. |
| Governing the AI tools themselves | A named owner for each tool | ISACA found that 64% of organizations have never run an AI-related incident response exercise. |
The work in the first three rows is repetitive and follows rules a machine can apply. I would hand it to AI without hesitation, since the alternative is skilled people collecting the same evidence by hand for every audit.
A control answer written once cannot show whether the control still works, and testing it again is exactly the work AI should take over. The reasons a point-in-time answer misses failures are set out in why point-in-time audits miss control failures. Our Security Posture Benchmark 2026, which we publish on 15 October, looks at the same problem in the security answers companies publish.
What DigitalXForce built AI to do, and what it leaves to people
We built our two AI engines around that split. AI JedAI analyzes: it reasons over control evidence and live telemetry, maps documents to controls and frameworks, scores and prioritizes risk, and recommends remediation mapped to framework requirements. XForce GPT writes: it produces the plain-language risk narratives and board-ready reports, and it generates policies, standards and plans.
In practice, our AI-Powered Risk Management and Automated GRC module tests each control against live data and maps the result once to 50+ compliance frameworks, and that is Continuous Control Assurance as we run it. Continuous Control Assurance (CCA) uses evidence, monitoring and validation to determine whether controls continue to operate as expected. Continuous Control Monitoring (CCM) monitors conditions, evidence and signals associated with controls. It is the capability within CCA that feeds those tests.
X-ROC, the XForce Risk Operations Center, handles real-time risk alerting, triage and escalation, and the board reports XForce GPT writes go to people who make the call. Neither engine accepts a risk. That decision belongs to a person, and it is recorded in the risk register of our Enterprise Risk Management module.
Governing the AI tools themselves is now part of the security job, since every AI tool a team uses also needs an owner, an entry in an inventory and a way to be switched off. We built AI TRiSCM™, the AI module of our Trust, Risk, Security and Compliance Management platform, for that work: it discovers AI assets, assesses models, copilots and agents, and maps them to the NIST AI RMF, ISO/IEC 42001 and the EU AI Act. The definition is in the glossary, the module details are on the AI TRiSCM and AI Risk Governance page, and the category is explained in what TRiSCM is.
Accountability stays with named people
Three kinds of decision stay with people, and the public frameworks and rules point the same way.
The first is judgment about what a finding means. A model can tell you a control failed. Weighing that failure against a product launch or a supplier’s contract terms needs context that mostly sits outside any system the model can read.
The second is risk acceptance. The NIST AI Risk Management Framework 1.0, released in January 2023, lists accepting a risk as one response option alongside mitigating, transferring and avoiding it (Manage 1.3). It also expects executive leadership to take responsibility for decisions about risks associated with AI system development and deployment (Govern 2.3). An AI system can prepare the file, and a person still has to sign it.
The third is governance and disclosure. Under the SEC cybersecurity disclosure rules adopted in July 2023, a company discloses a material cybersecurity incident on Form 8-K Item 1.05 within 4 business days of its materiality determination. That clock starts with a judgment, and the determination belongs to the people who will answer for it.
The data shows how often this is left unclear. ISC2 asked what happens when an AI recommendation leads to an incorrect outcome: 50% said human decision-makers are ultimately held accountable, 10% described ambiguity and 8% a lack of clear ownership. In ISACA’s March 2026 poll of more than 3,400 digital trust professionals, 20% did not know who would be responsible if an AI system caused harm, and only 36% said humans approve most AI-generated actions before they run. An organization that cannot say who owns an AI-assisted decision has a governance gap before it has a technology gap.
Practitioners say the same. ISC2’s July 2026 member paper on making automated security decisions defensible starts its list with governance and accountability, and it notes that accountability is often assumed instead of being defined.
What happens to entry-level roles
Entry-level roles are where the change is sharpest. ISC2 found that 56% of respondents think AI has somewhat or significantly reduced the need for entry-level positions over the past year, while 53% believe it is creating new types of entry-level roles. SANS found that among organizations with role changes, SOC and security analysts led the reductions at 32%, followed by threat intelligence analysts at 26% and incident responders at 22%. Stanford’s payroll data shows the same shape across the economy: the gap comes mainly from reduced hiring of young workers, and experienced workers show no comparable gap.
The tasks leaving the junior role, alert triage and log analysis, are the ones people used to learn the job on. If those tasks are automated, an organization has to decide on purpose how juniors will learn the judgment its senior roles depend on. ISC2 found that 62% do not believe AI has reduced the need for foundational cybersecurity skills, and I agree with them.
The skills a security team needs now
The shift is from volume to judgment. ISC2 found that 82% of respondents consider it very important to know when to trust AI outputs, and 80% when to override or ignore AI recommendations. ISACA found that 45% of respondents see LLM SecOps as a skill gap, 12 points higher than in 2025. The AI Workforce Consortium, which Cisco created with Accenture, Google, IBM, Microsoft and other companies, reported in August 2026 that cybersecurity job postings requiring AI skills doubled year over year across G7 countries, and Cisco’s summary puts the principle in five words: “People remain the final check.”
Headcount is no longer the headline. ISC2’s 2025 Cybersecurity Workforce Study of 16,029 practitioners dropped the workforce gap estimate that earlier editions carried, since its respondents ranked critical skills above more people. In the same study, 59% reported critical or significant skills needs. SANS found the same pattern: 60% of organizations now call skills the bigger problem, against 40% who cite headcount.
What I would do this quarter
- Write down every decision in your security and risk program that accepts a risk, and name the person who owns each one.
- Give AI the evidence collection and first-pass triage, and track how often a person overrides its output.
- Keep an inventory of the AI tools your security team uses, with an owner for each and a tested way to switch each one off.
- Run one AI-specific incident exercise this quarter, since ISACA found that only 8% of organizations do so regularly.
- Train analysts on when to trust an output and when to override it, because that is where their time is going.
Where DigitalXForce is not the answer
A small team with a few AI tools and no regulator asking about them does not need a platform for any of this. A named owner, a short policy and an inventory in a spreadsheet will serve it for now. When the control evidence outgrows the spreadsheet, DigitalXForce Lite is the full platform hosted in the cloud for any organization that prefers cloud hosting, with the same functionality. And no software, ours included, should be the one that accepts a risk.
Questions about AI and security teams
Will AI replace SOC analysts?
AI is taking over parts of the analyst’s day, starting with alert triage and log analysis, and SANS found that among organizations with role changes, SOC and security analyst roles led the reductions at 32%. The role is moving toward validating what the tools produce and deciding what to act on. A security operations center still needs people for those decisions, and why a traditional SOC is not enough on its own explains why DigitalXForce built X-ROC alongside it.
Which cybersecurity jobs are most at risk from AI?
The roles built around high-volume, repetitive work are the most exposed. SANS found that SOC and security analysts, threat intelligence analysts and incident responders led the AI-related reductions in 2026, at 32%, 26% and 22% of organizations with role changes. Roles built on judgment and governance are growing: among organizations adding roles, 32% added AI security engineers and 30% added AI governance analysts.
Will AI reduce entry-level cybersecurity jobs?
The data points both ways. ISC2 found that 56% of respondents think AI has reduced the need for entry-level positions, while 53% believe it is creating new types of entry-level roles. Stanford’s payroll research puts employment of workers aged 22 to 25 in AI-exposed occupations 19% below where it would otherwise be, driven mainly by reduced hiring rather than layoffs.
Is there still a cybersecurity skills gap if AI automates security work?
There is, and it is now a skills gap more than a headcount gap. ISC2’s 2025 Workforce Study of 16,029 practitioners did not publish a workforce gap estimate, and it found 59% reporting critical or significant skills needs. SANS found that 60% of organizations see skills as the bigger problem, against 40% who cite headcount, so AI has changed which skills are short without closing the gap.
How much time does AI save security analysts?
The most careful study I know of comes from Microsoft researchers, who found that adopting Security Copilot was associated with a 30.13% reduction in mean time to resolve incidents, using live operations data from 2024. It is vendor research on the vendor’s own product, and the authors say it shows association rather than proven cause. SANS found that 49% of organizations report reduced manual analysis time, which is a broader and less precise signal.
Can AI agents run a security operations center without people?
Not on the current evidence. The strongest agent in Carnegie Mellon’s TheAgentCompany benchmark completed 30% of 175 workplace tasks on its own, and ISACA found in March 2026 that 56% of digital trust professionals do not know how quickly they could halt an AI system during a security incident. An agent can run parts of the triage, and a person still has to own the decision and the off switch.
Who is accountable when an AI security tool gets it wrong?
A person should be, and that person should be named before the tool goes live. In the ISC2 survey, 50% said human decision-makers are ultimately held accountable, while 10% described ambiguity and 8% a lack of clear ownership. The NIST AI RMF expects executive leadership to take responsibility for decisions about AI risks (Govern 2.3).
Does DigitalXForce use AI to make risk decisions?
AI JedAI and XForce GPT do the preparation. AI JedAI analyzes control evidence, scores and prioritizes risk and recommends remediation, and XForce GPT writes the narratives and board reports. The decision to accept a risk belongs to a named person in your organization, and no tool should take it.
About this article
This article is Day 12 of the DigitalXForce Cybersecurity Awareness Month 2026 series, written by Margareta Petrovic. Every figure in it was read at its source on 25 September 2026.
Sources
- ISC2, “ISC2 Research: Rethinking AI’s Impact on Cybersecurity Roles”, published 14 July 2026, a survey of 856 cybersecurity professionals who use AI in their roles, conducted in May 2026. https://www.isc2.org/insights/2026/07/rethinking-ai-impact-on-cybersecurity-roles, read 25 September 2026.
- ISC2, “Member Paper on Building Accountability into AI Security Actions”, on the paper Making Automated Security Decisions Defensible, published 23 July 2026. https://www.isc2.org/Insights/2026/07/new-ISC2-member-resource-making-automated-security-decisions-defensible, read 25 September 2026.
- ISC2, “2025 ISC2 Cybersecurity Workforce Study”, published 4 December 2025, a survey of 16,029 practitioners and decision-makers in July and August 2025. https://www.isc2.org/Insights/2025/12/2025-ISC2-Cybersecurity-Workforce-Study, read 25 September 2026.
- SANS Institute and GIAC Certifications, 2026 Cybersecurity Workforce Research Report, published 11 March 2026, 947 respondents, with the SANS press announcement of its findings. https://www.sans.org/white-papers/2026-cybersecurity-workforce-research-report and https://www.sans.org/press/announcements/sans-research-cybersecurity-talent-shortage-narrative-wrong-real-crisis-what-your-team-doesnt-know-starting-ai, read 25 September 2026.
- ISACA, “Only 8 Percent of Organizations Conduct Regular AI-Specific Response Exercises, ISACA Research Finds”, published 22 September 2026, on the 2026 State of Cybersecurity survey of more than 1,800 cybersecurity professionals. https://www.isaca.org/about-us/newsroom/press-releases/2026/only-8-percent-of-organizations-global-enterprises-conduct-regular-ai-specific-response-exercises, read 25 September 2026.
- ISACA, “New ISACA Research: 56% of Digital Trust Pros Don’t Know How Fast They Could Shut Down AI After a Security Incident”, published 23 March 2026, a preview of the 2026 AI Pulse Poll of more than 3,400 digital trust professionals. https://www.isaca.org/about-us/newsroom/press-releases/2026/digital-trust-pros-dont-know-how-fast-they-could-shut-down-ai-after-a-security-incident, read 25 September 2026.
- World Economic Forum, Global Cybersecurity Outlook 2026, January 2026, a survey of 804 qualified participants from 92 countries run from 25 August to 1 October 2025. https://reports.weforum.org/docs/WEF_Global_Cybersecurity_Outlook_2026.pdf, read 25 September 2026.
- Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, Stanford Digital Economy Lab, “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence”, working paper revised 12 August 2026, using ADP payroll data through June 2026. https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/, read 25 September 2026.
- Frank F. Xu, Graham Neubig and others, Carnegie Mellon University, “TheAgentCompany: Benchmarking LLM Agents on Consequential Real World Tasks”, arXiv 2412.14161, version 3 revised 10 September 2025. https://arxiv.org/abs/2412.14161, read 25 September 2026.
- Ayush Chopra, Ramesh Raskar, Prasanna Balaprakash and others, Massachusetts Institute of Technology and Oak Ridge National Laboratory, “The Iceberg Index: Measuring Skills-centered Exposure in the AI Economy”, arXiv 2510.25137, 2025. https://arxiv.org/abs/2510.25137, read 25 September 2026.
- James Bono, Justin Grana and Alec Xu, Microsoft Corporation, “Generative AI and Security Operations Center Productivity: Evidence from Live Operations”, arXiv 2411.03116, November 2024. This is vendor research. https://arxiv.org/abs/2411.03116, read 25 September 2026.
- Cisco, “AI Workforce Consortium: Building a cybersecurity workforce ready for what’s next”, Cisco Blogs, published 20 August 2026. https://blogs.cisco.com/our-corporate-purpose/ai-workforce-consortium-building-a-cybersecurity-workforce-ready-for-whats-next, read 25 September 2026.
- NIST, AI Risk Management Framework (AI RMF 1.0, NIST AI 100-1), released 26 January 2023, intended for voluntary use. https://www.nist.gov/itl/ai-risk-management-framework, read 25 September 2026.
- NIST Trustworthy and Responsible AI Resource Center, AI RMF Core, section 5: Govern 2.3, Govern 3.2, Map 3.5 and Manage 1.3. https://airc.nist.gov/airmf-resources/airmf/5-sec-core/, read 25 September 2026.
What this looks like in practice.
Reading about continuous evidence is one thing. Watching a control get tested against live data from your own stack is another. A 30 minute walkthrough on your frameworks shows the difference.
