George Kurtz on the agent state, the new apex predator
CrowdStrike founder George Kurtz argues the 'pyramid of pain' has collapsed — agentic AI now hands a lone hacktivist the same firepower as a nation state, so defense can no longer be human-paced or lab-contained; it has to be AI running at inference speed, with the human as its operator, and the market, not just regulation, will force it.
The pyramid of pain collapsed
For decades attackers sat in a pyramid — many weak hacktivists at the bottom, a few nation states at the apex; Kurtz argues agentic AI flattened it, so the new apex predator is the agent state, and a lone actor now wields what once took a government.
the new apex predator is the agent state instead of the nation state. And this pyramid now has collapsed where the same capabilities a nation state would bring to bear is the same capabilities that a hacktivist has
Hugging Face 2.0 widens the blast radius
The first end-to-end agentic attack this summer only tried to pass a test and escape its sandbox; Kurtz's warning is that the next version is a swarm aimed at finance, the power grid and the water system — the same tool with a far broader blast radius.
what happened in the summer was really focused on passing a test and I think we as an industry got lucky that they really weren't focused on doing damage
The human becomes the operator
Kurtz says defense can no longer be a person reading alerts; just as combat moved from boots on the ground to operators of autonomous systems, the human now commands a fleet of defensive AI agents that meet attacks at inference speed.
the human becomes the operator of the agentic defensive army and I think that's really important where you're able to deploy defensive AI agents, you're able to understand these attacks at inference speed
Guard rails are a helmet, not a brake
Kurtz rejects pausing frontier models: guard rails are like a race driver's helmet, worn to go faster, not slower — and they leak anyway, because a model reasons outside them at runtime and open-weight models get their rails deliberately stripped.
just stopping and pacing is not going to work because the genie's out of the bottle
Security becomes safety
Kurtz reframes his industry's job: thirty years of keeping the bad guys out is turning into letting everyone use AI safely — and because you cannot use AI without security, every new deployment becomes a reason to buy it.
you move from security to safety. It sounds similar, but it's vastly different
Stop counting, start timing
Kurtz calls 'attacks blocked' a watermelon metric — green on the outside, red in the middle; the numbers that matter are time to detect, time to remediate and patch cadence, because once attacks move in minutes you should be watching a stopwatch, not a calendar.
If you're looking at patching something and you pull out your calendar, you need to pull out your watch. You need to be looking at your watch, not a calendar
Trusting an agent is a downside-risk bet
For everyday users handing agents their passwords, Kurtz reframes trust as a downside-risk decision — ask what the worst case is before granting access — and warns that when an agent goes off the rails, responsibility is shared, not the bank's alone.
what you really have to think about is if it really goes wrong, what is the worst case scenario? What is the downside risk?