AI agents are taking the first pass at a growing share of security alerts. An agent gathers the evidence, compares the case with what it has seen before, and writes up a verdict with a summary attached. Someone still has to assess whether that verdict is correct. With the gathering and matching already done, that assessment is the part of the case left to a person, so judgment now matters from the first day of a security career. Routine cases were where that judgment formed. As agents handle more of those cases, fewer of them reach new practitioners, so the volume of practice that built judgment for earlier practitioners is harder to come by.
Judgment has always formed through practice and guidance, by working real cases while an experienced practitioner points out what you missed. That guidance was never easy to come by, because experienced practitioners carry their own caseloads and have limited time to walk anyone through every alert. The same agents that are taking over the routine work can help close that gap. Research on how people learn with AI shows that the design of the help decides whether it builds skill or erodes it. Judgment is becoming the entry-level skill. AI agents can help build it when teams use them to teach as deliberately as they use them to detect.
What the Agent’s Verdict Leaves Open
For as long as security teams have worked alert queues, a new practitioner proved themselves by volume: working through cases quickly, closing the routine ones, and passing the rest to someone more experienced. Speed was what a manager could measure, so speed was what a first-year review tended to praise. Agents now take the first pass at more of those cases before any person sees them, working the routine ones faster than a person can. The entry-level work is being hollowed, not eliminated, because the seat stays while the fast part of what filled it moves to software. What remains is the slower question of whether a verdict is right when the evidence is incomplete, a question that has never been fast for anyone.
Picture an agent’s summary that says an account signed in from an unfamiliar country and that the sign-in was harmless, because the user is traveling. Someone has to know whether the travel record the agent relied on is current. Someone has to know whether that user’s access makes the sign-in matter more than usual, along with whether the same pattern appeared on three other accounts that morning. Each of those questions takes context the agent may not have, as well as the judgment to recognize when a confident answer is wrong. That judgment, assessing whether an agent got it right and saying so when it did not, is what the entry-level job increasingly requires.
The Skills Gaining Value
Four practitioner skills gain value when agents take the first look. The first is business context, meaning knowledge of what a system does for the organization, who depends on it, and what an hour without it costs. That knowledge turns a technically correct verdict into the right decision for the people who rely on the system. The second is adjudication, the plain skill of deciding whether a verdict is right and saying so in writing when it is wrong. When an agent labels an alert harmless and the label turns out to be wrong, a person still answers for it.
The other two concern the machinery. The third is detection engineering, which is the work of writing and tuning the rules that decide what counts as suspicious in the first place. When agents act on those rules at machine speed, a badly written rule produces bad verdicts at machine speed too, so the person who understands why a rule fires holds more of the outcome than before. The fourth is agent orchestration, which means directing several automated tools through a piece of work and checking each hand-off between them. All four join an engineer’s standard of evidence to a technologist’s judgment about consequence.
Judgment is becoming the entry-level skill. AI agents can help build it when teams use them to teach as deliberately as they use them to detect.
Thinking Like a Technologist
I’m a technologist, but I am not an engineer by training. I came to cybersecurity after years as a US Army intelligence officer. I developed a habit of reading a system from the outside: what it does, what someone intends to do with it, and what happens when it fails. That habit is rising in value now, because when the machine takes the first look, someone still has to decide whether the machine is right. My judgment came the slow way, one case at a time, from people who showed me what they saw in cases we both looked at.
The difference between the two roles matters for anyone deciding what to learn. An engineer builds a system and proves that it does what it was built to do. A technologist works out what a system is good for, what could go wrong with it, and what that would cost the people who rely on it. Security teams increasingly need the second kind of thinking from people early in their careers, because an agent can produce the technical answer while a person still has to weigh what it means. That thinking rests on real knowledge of how systems work. It develops by working real cases on top of that technical depth.
The Design of the Help Decides
How AI helps decides whether people learn from it. When a company gave 5,172 customer support agents an AI assistant built from the habits of its best agents, the least-skilled agents gained the most, while agents who had used it kept some gains during outages. Math students in Turkey who practiced with a plain chatbot did 48 percent better while they had it and 17 percent worse once it was taken away, although a hint-giving version built with teachers’ input largely avoided that harm. Developers learning an unfamiliar programming library scored 17 percent lower on a quiz when they had an assistant, though those who used it to ask conceptual questions scored well. Help builds skill when it leaves the learner doing the work and erodes skill when it does the work for them. As the developer study’s authors put it, “AI-enhanced productivity is not a shortcut to competence.”
Agents That Teach
Judgment has traditionally come from a real case worked together with someone experienced, with something at stake, which is an occasion for learning. Experienced practitioners carry their own caseloads, so many have had the week where the only guidance they could give was “close it, it’s harmless, next.” A team’s metrics count the cases closed, so the teaching that never happened leaves no trace in them. Used deliberately, the agents a team already runs can supply part of that guidance on every case, at the moment a new practitioner needs it. A new practitioner can be asked for a verdict before seeing the agent’s, can read the evidence the agent relied on, and can ask it to explain why a case matters.
Agents teach only when the teaching is designed in, because the plain version of an AI tool, used to finish the work, is the version that left students worse off once it was gone. The people who decide how agents are used, and how new practitioners work alongside them, are the ones who choose between those two versions. Designed well, the arrangement puts experienced practitioners’ limited time where it counts most, on the cases where a new practitioner and an agent disagree. Those disagreements are where a new practitioner’s reasoning is most visible and easiest to correct. The agent handles the volume, leaving the experienced practitioner the teaching moments the volume produces.
A Harder Start, and a Way Through
The hardest part of this shift falls on the new practitioner. Adjudication, business context, and the confidence to overrule a confident machine are skills experienced practitioners used to reach after years of routine cases. The work that once came first is moving to software, so new practitioners are being asked to start near where their predecessors finished. That is a real increase in difficulty. It also means the most tempting shortcut, letting an AI tool finish the work so the numbers look good, is the one the evidence says costs the judgment the job now pays for. Starting there is still possible, because judgment is built one real case at a time, with an experienced practitioner when one is available and with an AI tool set up to question your reasoning on every case in between.
Help builds skill when it leaves the learner doing the work and erodes skill when it does the work for them.
The Next Step
If you are trying to get in, you can practice judgment before you have the job. Take a public breach report or a government security advisory, read only the timeline and the evidence, and write down what you think happened and what the affected organization stood to lose. Then read the conclusion and note where your reasoning went wrong, which is where adjudication and business context begin. Free defensive-security challenge labs give you realistic logs to work through the same way. When you use a general-purpose AI chat assistant, tell it to hold back the answer until you have given yours, then ask it to critique your reasoning and explain what you missed.
If your team is piloting AI agents for triage, use them to teach as well as to close cases. For one routine alert type the agents already handle, have each new practitioner write their own verdict, with the evidence behind it, before reading the agent’s. Once a week, have an experienced practitioner spend half an hour with each new practitioner on the cases where the two verdicts disagreed. Count the adjudications reviewed alongside the cases closed, because a review that rewards speed alone trains new people in the one skill the field is handing to software. Judgment is becoming the entry-level skill. AI agents can help build it when teams use them to teach as deliberately as they use them to detect.
A note on where I stand: I spent years running security inside large financial institutions. I also sit on the advisory board of a company building autonomous security tools, which is the category whose work this issue describes. Nothing non-public, from anywhere I have worked, goes into what I write here.
Next: where the new first jobs are and where these skills get practiced.