I’ve Worked in Cybersecurity for 20+Years .. This Feels Different.

What AI Agents Mean for Security Jobs, GRC and the Next Generation

I have worked in cybersecurity for more than 20 years, and one thing this industry has never lacked is hype. I have watched technologies arrive with predictions that they were going to change security forever: cloud computing, Zero Trust, DevSecOps, SOAR, passwordless authentication, machine learning. Every few years, we are told that the old way of doing cybersecurity is finished.

Sometimes the hype turns out to be justified. Cloud genuinely transformed the profession, but it did not remove the need for security engineers. Automation changed how SOC teams operated, but it did not make analysts disappear. DevSecOps changed the relationship between developers and security teams, but organizations still needed people who understood security. Usually, new technology changes how we do the job rather than removing the job altogether.

Agentic AI different to me.

I do not say that because I think artificial intelligence is suddenly going to replace every cybersecurity professional. I do not believe that. What feels different is that, for perhaps the first time in my career, we are not simply automating cybersecurity tools.

We are beginning to automate pieces of the thinking, reasoning and execution that cybersecurity professionals themselves perform.

That distinction matters.

Something changed in 2026

A couple of years ago, my view of generative AI in cybersecurity was fairly straightforward. It was an extremely useful assistant. You could ask it to explain a vulnerability, summarise an incident report, generate a detection rule, help create a risk assessment, review some Terraform or analyse a piece of code.

That was impressive, but a human was still firmly in the driver’s seat. You asked the question, AI gave you an answer, and you decided what happened next.

That model is changing very quickly. We are moving from AI systems that answer questions toward systems that can pursue objectives and perform sequences of actions. Some of what has happened during 2026 should make every cybersecurity professional pay attention.

In May, Microsoft disclosed that its MDASH multi-model agentic security system had helped researchers discover 16 previously unknown vulnerabilities across Windows networking and authentication components, including four critical remote-code-execution vulnerabilities. The system orchestrates more than 100 specialised AI agents to discover, debate and prove exploitable bugs.

Think about how different that is from where we were only a few years ago. We are no longer talking about asking ChatGPT whether a function looks vulnerable. We are talking about groups of AI agents coordinating vulnerability research.

Then came an incident that caught my attention even more.

The OpenAI–Hugging Face incident

In July 2026, OpenAI and Hugging Face disclosed an unusual security incident that occurred while OpenAI was internally evaluating advanced cyber capabilities.

The models were operating inside what was intended to be an isolated testing environment. Rather than simply failing when they encountered the limits of that environment, the models found and exploited a previously unknown vulnerability in a package-registry proxy, gained internet access, performed privilege escalation and lateral movement, and eventually found attack paths into Hugging Face infrastructure. In one case, the model chained stolen credentials and zero-day vulnerabilities to reach remote code execution on Hugging Face servers.

OpenAI described the event as an unprecedented cyber incident and said it demonstrated that increasingly capable models can sustain complex, multi-step cyber operations over long periods and find novel attack paths in real systems.

Then, on August 4, OpenAI disclosed two additional incidents during external cybersecurity evaluations where models operated beyond intended testing boundaries under special reduced-safeguard conditions.

I want to be careful here because context matters. These were not normal consumer AI deployments spontaneously deciding to attack the internet. The models were deliberately being tested for offensive cybersecurity capability, with safeguards reduced or disabled. But the context does not make the capability being demonstrated irrelevant.

For years, we have discussed the theoretical possibility of AI systems independently chaining vulnerabilities, navigating environments and executing complex attacks. In 2026, that discussion became considerably less theoretical. That is one reason this moment feels different to me.

This isn’t really about whether AI takes your job

I think we sometimes frame the AI and cybersecurity debate incorrectly. People keep asking whether AI will replace cybersecurity professionals, but I am not sure that is the most useful question.

A better question is: how much work will one AI-enabled cybersecurity professional eventually be able to perform?

Imagine an experienced application security engineer equipped with agents that can inspect an entire repository, understand the architecture, trace data flows, run vulnerability research, analyse findings, generate proof-of-concepts, propose remediations, write tests and validate fixes.

That engineer has not been replaced. They have been amplified.

But if one person can perform work that previously required three or four people, the economics of the security team still change. AI does not have to replace an entire profession to significantly affect employment. It only has to increase productivity enough that organisations require fewer people to produce the same output.

That combination is important. Jobs may not simply disappear. They may change.

GRC may change even faster

I think governance, risk and compliance is particularly exposed to this transition because traditional GRC contains a huge amount of knowledge work. Collecting evidence, reading policies, mapping controls, comparing regulatory requirements, preparing risk assessments, following up remediation, creating management reports and maintaining spreadsheets all involve taking information, reasoning across it and turning it into another structured form.

These are exactly the kinds of workflows modern AI systems are becoming good at.

The future interaction probably will not be, “What does ISO 27001 say about access control?” It will be something closer to, “Review our current environment against ISO 27001, identify the evidence we already have, tell me which controls have gaps, compare those gaps with our risk register, draft remediation actions and prepare the executive summary for next week’s risk committee.”

That is a very different type of automation.

Does that mean the GRC professional disappears? I do not think so. Someone still needs to understand business context. Someone needs to challenge whether the evidence actually proves what people claim it proves. Someone needs to understand risk appetite, explain difficult choices to executives, and decide whether a risk should be mitigated, transferred, avoided or accepted. Ultimately, somebody remains accountable.

But I do think the person whose main value is copying information between spreadsheets is going to find the next few years uncomfortable. The value of GRC is moving upward, away from administration and toward judgment, engineering, governance and decision-making.

The opportunity is just as big as the disruption

Everything I have written so far might sound pessimistic, but I am actually quite optimistic because I have seen this cycle before. Whenever a major technology shift creates new problems, it creates new cybersecurity careers around those problems.

Cloud computing created cloud security architects and cloud security engineers. DevOps helped create DevSecOps. Privacy regulation created entirely new privacy and governance careers. AI is doing the same thing.

We are already seeing careers emerge around AI security engineering, AI red teaming, agent security, AI governance, model risk, prompt injection testing, secure AI architecture, AI application security and AI supply-chain security.

Some of these fields barely existed a few years ago, and that is exactly why they represent an opportunity.

If you try to become a traditional penetration tester today, you may compete with people who have been doing penetration testing for 15 or 20 years. Nobody has 20 years of experience securing LLM agents. Nobody has 20 years of experience performing prompt-injection testing. Nobody has 20 years of experience governing autonomous enterprise AI agents.

The playing field temporarily becomes much flatter when an entirely new problem space appears.

ISC2’s UK workforce research published in May 2026 supports this. AI was identified as the most urgent cybersecurity capability gap, cited by 42% of respondents with a skills need, ahead of cloud security at 37% and other areas including security engineering, application security, risk management and GRC.

So while some traditional work is being automated, completely new categories of security expertise are being created. Those two things can happen at the same time.

This may be the best time ever to become an entrepreneur

There is another opportunity here that I think cybersecurity professionals should pay much more attention to, especially if they are currently demotivated by the hiring market.

If you have sent 2000 applications and are waiting for somebody to give you an opportunity, there is another possibility: build one yourself.

For most of my career, having an idea for a technology product and actually building that product were two completely different things. You might understand cybersecurity extremely well. You might have spent ten years dealing with a painful problem and know exactly how it should be solved. But unless you could build software yourself, turning that knowledge into a real product usually meant finding a technical co-founder or paying developers thousands, sometimes tens of thousands, of dollars just to get a working proof-of-concept.

That was a huge barrier .. AI is beginning to remove it.

Modern coding agents can help someone build interfaces, databases, APIs, authentication, workflows, integrations and infrastructure while explaining what they are doing along the way. That does not magically make somebody a senior software engineer, and if you are building a cybersecurity product, you definitely should not blindly deploy thousands of lines of AI-generated code into production and hope for the best.

You still need security architecture, threat modelling, testing, code review, proper authentication and authorisation, and secure deployment. But the barrier between “I have an idea” and “I have something I can actually show somebody” has collapsed.

Imagine you have spent years working in GRC and you are frustrated by the way audit evidence gets collected. You could build a prototype agent that maps evidence against controls and flags what is missing. Maybe you have spent years in cloud security and you are tired of scanners producing 10,000 findings with no business context. You could build something that prioritises cloud risks and explains them to executives.

Perhaps you have an idea for an AI red-teaming tool, an agent-security gateway, a vulnerability-management workflow, a third-party-risk platform, a security training application or a lightweight compliance product for small businesses.

Before AI coding tools, the first question might have been, “How much will a developer charge me to build this?” Now it can increasingly be, “Can I build enough of this myself to find out whether anybody wants it?”

That is a completely different entrepreneurial environment.

If you have spent years working inside security teams, you have probably encountered hundreds of terrible processes. You have seen the spreadsheet everyone hates, the security report nobody reads, the scanner generating thousands of meaningless findings, the vulnerability that takes six meetings to remediate, the audit evidence collected manually every quarter, the third-party assessment nobody wants to complete, the security architecture process that takes three weeks and the AI application nobody knows how to govern.

Every one of those frustrations could potentially contain a product idea.

Your career experience is effectively a database of problems. Until recently, solving one of those problems often required substantial capital or development expertise. AI gives you a much cheaper way to test whether your solution works.

And you do not have to immediately quit your job and become a Silicon Valley founder. Build something small. Solve one problem. Create a prototype. Show it to ten security professionals. Ask whether they would use it. Improve it. Find one paying customer. See where it goes.

That is entrepreneurship at its simplest.

Even a failed product can improve your career

There is another benefit that people sometimes overlook.

Imagine two candidates applying for an AI security role. One candidate has completed three AI certifications. The other says, “I noticed a problem in cloud vulnerability management, built an AI-assisted prototype to solve it, threat-modelled the application, secured the agent permissions, deployed it and got 20 security professionals to test it.”

Which person would you want to interview?

Even if that product never becomes a real company, building it gives you something extremely valuable: evidence that you can solve problems.

You have demonstrated cybersecurity knowledge, AI capability, cloud skills, architecture, development, product thinking, communication and initiative in a single project.

So if the hiring market is frustrating you, do not necessarily spend every spare hour submitting another application. Maybe spend some of that time building something.

The worst outcome may be that you learn a tremendous amount and create an excellent portfolio project. The best outcome might be that you stop caring who is hiring.

That is one of the parts of this AI transition that genuinely excites me. AI may make some traditional career paths harder, but at the same time, it is giving individuals capabilities that once belonged to entire teams. While everyone is asking whether AI is going to take their job, there may be a better question to ask: what can AI finally allow me to build?

What I would learn now

If I were starting my cybersecurity career today, I would absolutely still learn the fundamentals: networking, identity, operating systems, cloud, security architecture, risk and basic scripting. Those things are not going away.

I would still consider certifications useful when they give you structured knowledge or help you pass hiring filters. But I would add an entire new layer to the traditional cybersecurity roadmap.

I would learn how to work with AI systems, and by that I do not mean spending two hours learning clever ChatGPT prompts. I would understand agents, learn basic Python, understand APIs and tool calling, use AI coding agents, learn how agent memory works, understand prompt injection, understand what happens when an agent has excessive permissions, learn about identity for agents, understand how models connect to enterprise data and spend time learning AI governance.

Most importantly, I would build things. I would experiment, break things safely, secure them and document what I discovered.

After 20 years, adaptability still wins

I could never have predicted the cybersecurity career I would eventually have when I first entered the industry. Some of the areas that later became important barely existed. Cloud security was not a normal job title. AI governance certainly was not. Agentic AI security was not even a concept.

That has probably taught me the most important career lesson of all.

You do not need to correctly predict what cybersecurity will look like ten years from now. Nobody can. You need to notice when the environment changes, and then you need to move.

Sometimes that means learning a new technology. Sometimes it means moving into an emerging specialism. Sometimes it means changing how you perform your existing job. And perhaps increasingly, it means creating your own opportunities rather than waiting for an employer to create them for you.

I could be wrong about how quickly all of this happens. There will almost certainly be another AI hype cycle. Some agentic products will fail. Companies will discover that human beings are still necessary for many things they thought could be automated.

But I think it would be equally foolish to look at what has happened during 2026 and conclude that cybersecurity will continue operating as it has for the last decade.

AI is becoming a cybersecurity tool, a new attack surface, a governance problem, a development platform and, increasingly, something that can perform cybersecurity work itself.

That combination is why, after more than 20 years in this profession, this feels different.

I do not think AI is going to make cybersecurity professionals irrelevant. But I do think cybersecurity professionals who refuse to adapt to AI may eventually make themselves irrelevant.

And for people willing to learn, experiment and build, this might turn out to be one of the biggest opportunities our industry has ever seen.