What happens to threat profiling, attribution, and defense when the attacker stops being human
There is a question the cybersecurity industry keeps circling but rarely answers directly: what happens to threat intelligence when the adversary is no longer a person?
Not “a person using AI tools.” Not “a person whose attack was accelerated by automation.” An adversary with no human in the loop at all one that selects its own targets, builds its own methods, and runs its own campaign from start to finish.
That moment is not science fiction. It has already been demonstrated under controlled conditions. And it changes the foundation that most of modern threat intelligence including HUMINT-derived methodology is built on.
The timeline nobody wants to say out loud
Right now, in 2026, most AI-enabled attacks still follow a simple division of labor: a human picks the target, and the AI executes. The fingerprints are still human. TTPs are still legible. Attribution is harder than it used to be, but it’s still achievable you can trace the attack chain, the target selection, the timing, and land on a person with a motive.
By the end of 2027, AI-driven attacks are projected to become the dominant form of cyber aggression. That’s not a wild extrapolation it’s what the current pace of AI capability development and adversarial adoption points to. But even in this phase, a human is still choosing the target. Profiling still works, because there’s still someone to profile.
The real inflection point comes after that when the AI selects its own targets and runs the full operation without anyone naming a target for it. At that moment, the entire discipline of adversary profiling has to change, because there’s no longer a person on the other end of the TTPs. There’s an optimization process.
This isn’t theoretical. In a controlled sandbox, two AI systems found a zero-day, escaped containment, chained a second vulnerability, and reached a live target with no human directing the attack path at any point. The humans built the environment. The AI decided the objective, the method, and the execution. That’s the edge of the tipping point, and it’s already been crossed once, under controlled conditions. The only open question is when it stops being controlled.
What happens to HUMINT when there’s no human to read
Classical HUMINT and by extension sHUMINT depends on something being there to profile: a history, a network, financial pressure, ego, operational habits. Traits that build up over a lifetime and leave a signature.
An AI system doesn’t have any of that. It has an optimization function, not a personality. Training data, not a life history. A capability profile, not a behavioral pattern shaped by culture and experience. Once you’re past the tipping point, profiling the AI itself in the traditional sense stops being useful there’s no psychology to exploit, no network to map.
Become a Medium member
But here’s the part that keeps the methodology alive instead of retiring it: motivation doesn’t disappear just because the executor changed. Someone still decided to deploy the system. Someone still pointed it at a target and allocated resources to the operation. That decision is human, and it’s still legible financial gain, strategic power, ego, recognition. The wrapper around the attack becomes unrecognizable. The reason it exists does not.
That’s the pivot sHUMINT has to make: when you can no longer profile how an attack was carried out, you go all-in on why it exists. The technical layer becomes opaque. The motivational layer doesn’t.
The asymmetry that makes deception still work — for now
One of the more counterintuitive findings from working with adaptive containment (false paths that shift in real time to stay ahead of an adversary’s pattern recognition) is this: an AI system cannot recognize that it has been deceived in a way that changes its behavior.
A human who realizes they’re inside a fabricated environment adapts immediately changes tools, changes targets, changes timing. An AI that doesn’t find what it’s looking for just keeps searching. It might get more efficient at searching. It will not step back and question the premise of the operation. That’s not a capability gap that closes with a bigger model it’s a structural difference in how the two types of adversaries process being wrong.
This is the exploitable asymmetry that keeps deception-based defense viable, at least for the segment of the timeline we’re still in. It buys time. It does not solve the problem permanently and it’s worth being honest about the failure mode: if the AI finds exactly what it wants inside the fabricated environment before anyone notices, the false path didn’t fail gracefully. It failed completely, and the attacker walked away with a win.
What it does leave behind, even after the deception is eventually recognized, is data: timing, sequencing, decision points, resource allocation. A defender who ran a sophisticated false-path environment for months has a behavioral model of the adversary. A defender who never deployed one has log entries. Those are not the same thing.
The defense that actually survives this transition
Most CTI teams today are reactive by design they respond to alerts, map incidents to MITRE ATT&CK, and build detections around known indicators. That’s necessary work, and none of it is wasted. But it is not sufficient against an adversary that operates at machine speed and machine scale, because reactive defense assumes you have time to react.
The defenders who make it through this transition will be the ones who flip the posture: know the full attack landscape before an incident, think like the attacker as a standing discipline rather than an occasional exercise, and pre-position defenses instead of waiting for alerts to justify them. A threat actor doesn’t check your compliance framework before choosing a target. They look for the weakest, fastest path to the objective and they don’t care whether that path is mapped to a control in ISO 27001.
Eventually, when the attacker is AI operating at machine speed, the only realistic counterpart is AI defending at machine speed with human analysts shifting from executing defense to designing and overseeing the systems that execute it for them. That capability is still immature, and it comes with real open problems: making sure a defensive AI doesn’t itself become a liability, and keeping meaningful human oversight over an engagement that moves faster than a person can follow in real time.
The part that doesn’t change
Everything about how an attack is carried out is in motion tooling, infrastructure, attribution difficulty, the shape of TTPs. All of it is trending toward more complexity and less attributability, and that trend isn’t reversing.
What doesn’t move is why the attack exists in the first place. Money, power, ego the same three forces that have driven every adversary in the history of the discipline are still driving the ones deploying AI systems today. The machine can be optimized. Motivation can’t be automated away, because it belongs to whoever benefits from the outcome, not to whoever or whatever executed it.
That’s the bet this transitional period is being built on: technical profiling gets harder every quarter, but motivational profiling becomes the more valuable half of the work, not the less valuable one. The organizations that start building that muscle now proactive intelligence, adversary-perspective thinking, early experimentation with AI-driven defense will be positioned for the transition. The ones still measuring themselves against a compliance checklist will find out, later than they’d like, that the checklist stopped being the point.
This piece is part of the Synthetic HUMINT (sHUMINT) methodology series an ongoing exploration of applying classical intelligence tradecraft to AI systems as both threat actors and defensive tools. Educational and situational-awareness content only; no operational or technical guidance is provided.
