There is a fear that comes up every time AI applied to security is discussed: that autonomous systems will end up deciding on their own and leave the human team out of the equation. I think that framing is backward. The real risk is not that AI turns against the organization, but that someone, inside or outside, uses it to attack it. Automation has no will of its own. It executes the intent of whoever directs it.
And within that whole landscape of faster attacks, exposure surfaces that keep growing and constant pressure to cut costs, there is one idea worth rescuing. No matter how much you automate, defense still depends on people. And that dependence is not a temporary limitation that the next version of the model will solve. It is structural.
What automation cannot replicate
A simple question dismantles much of the fatalism: have you, as an analyst, ever been the most capable person on your team? Almost never. And yet you still add value. Raw processing power was never the only thing that made a security professional useful. An engine can correlate millions of events per second, but that does not make it the one who decides what matters.
Context and judgment cannot be manufactured
A system can classify an alert as critical or benign. What it does not know is that the server hosts the month-end accounting close, or that the anomalous behavior coincides with a planned migration, or that the vendor has been in an access renewal process for three weeks. Business context, the kind that builds up over the years and is almost never documented in a log, is an asset the machine does not have and that the organization needs in order not to drown in false positives. Or worse, so as not to dismiss the one signal that actually mattered.
Something quite practical follows from this. The roles that survive automation are the ones that rely on judgment and relationships: the analyst who interprets the alert in its context, the leader who translates technical risk into a business decision, the team that manages a crisis when the automated playbook falls short.
The analyst as an interface, not a replaceable part
It is worth thinking of today's security professional as someone who borrows capacity from AI and turns it into defensible decisions. The model contributes speed and scale. The person contributes judgment, accountability and credibility before management, auditors or a regulator. That is why it is not enough to buy intelligence directly from the machine: someone has to answer for the decision, and that someone is still human.
Trust as the base currency
A large part of security is not technical, it is relational. Social engineering exploits human relationships, insider threats arise from them and incident response rests on trust between teams under pressure. Communicating a risk to a non-technical executive, convincing a business unit to apply a patch that breaks its workflow, keeping calm in a war room at three in the morning. None of that gets automated. Trust remains the currency you operate with when things get serious.
Computing power is not everything
An image helps. The most powerful system in the lab does not last five minutes against a real adversary who changes tactics, lies, improvises and exploits exactly what was not in the training data. Organizations do not defend themselves by having the biggest engine, but through coordination among people, the tacit knowledge that circulates among them and the ability to adapt to what no tool anticipated. Reducing security to “more compute” is convenient, but incomplete.
Alignment looks more like parenting than control
This is where the dependence becomes mutual. Talking about controlling an autonomous defense system is partly an illusion, because we rarely fully control a complex environment. The useful question is not whether the system will be faster than us, because it will be. The question is whether it will act according to the organization's correct priorities.
And that lives less in the code than in the data and the policies with which we train and tune the system. That is why models and automations sometimes show behaviors that even their own operators cannot fully explain: a rule that blocks too much, an agent that escalates privileges because it misread a general objective, an automated response that fires where it should not. It is almost never an explicit instruction. It is something the system infers from a broad principle. Automation depends on what we teach it, and we depend on that learning having been good, auditable and reversible.
The rest of the landscape, in brief
A few more ideas that frame the current moment.
The gap between what is visible and what is real. What generates the most noise, the spectacular demos and the viral incidents, tends to be overblown and not very effective. What really changes the game happens quietly: systems that iterate on themselves and adversaries that automate reconnaissance, exploitation and evasion continuously. Sooner or later they find the gap.
Disruption arrives at the entry level. The most routine work (tier 1 triage, basic correlation, indicator enrichment) is the first to be automated. Roles that require more judgment and the most manual tasks hold out considerably longer. The organizational risk is not being left without a SOC, but emptying out the rung on which future experts were trained.
Offensive automation weighs more than unemployment. Attacks have become cheap and scalable. Agents capable of chaining together complete phases of an intrusion lower the barrier to entry and shorten timelines. Defense has to accept that asymmetry.
Choosing a vendor is also a security decision. Not only for technical capability, but also for how the people who build the tool behave: what they do with your data, how much transparency they offer and what they are willing to sacrifice to do things right.
What to do with all this?
Four concrete things for teams and professionals.
- Learn to use AI in defense. It is not the enemy. The better you integrate it into detection, response and analysis, the more capable your team will be. And use it to sharpen your judgment, not to switch it off: instead of asking it to resolve the alert, ask it for the reasoning, the alternative hypotheses and what still needs to be verified.
- Prepare for a hybrid SOC. Understand the workings of agents and automations, where it makes sense to keep a human in the loop and how to design reversible, auditable controls.
- Double down on the human side. Contextual judgment, risk communication and crisis management do not get automated. Make them your differentiator.
- Do not let go of governance. The pressure to move fast pushes people to loosen the reins. The alternative is to build automation that is traceable, explainable and aligned with the organization's real priorities, a principle that frameworks such as the NIST AI Risk Management Framework already formalize, and not just with the quarter's metric.
Conclusion
Automation is very soon going to be better than us at almost any measurable security task: correlating, classifying, responding in milliseconds. But measurable is not everything. Our dependence on the machine grows every day. Its dependence on us, on our context, on our judgment, on the trust that holds teams together and on the values we train it with, remains structural. The future of cybersecurity does not depend on the machine surpassing us, but on the kind of professionals we decide to be while we build it.
