No Priors

George Kurtz on Fighting AI Attacks With Better AI

George Kurtz· Founder and CEO at CrowdStrike
·~45 min·English·No Priors
AgentsAI SafetyLLMAI CompanyAI Infrastructure
TL;DR

CrowdStrike founder George Kurtz explains how AI flattened the attacker pyramid and collapsed breakout time to seconds, and why the only defense is arming defenders with equal-caliber AI trained on data no lab can scrape.

01The New Threat

The New Apex Predator Is the Agent State

AI flattened the attacker pyramid — the least-skilled hacktivist now wields capability that only a nation-state had a few years ago.

The new apex predator is the agent state. The same level of capability that a nation state would have just a few years ago, essentially a hacktivist would have today.

— George Kurtz, No Priors
Key Insight
Defense used to triage by attacker tier: ignore the noisy low-skill base, watch the rare top. With the pyramid flat, attacker sophistication no longer tells you who to worry about, so the sheer volume and speed of attacks matter more than who is behind them.
02The New Attack

It's Not Malware Anymore — It's Prompts

Attacks no longer have a fixed signature: each run is generated from a prompt, so there is no malware file to match and block.

it's not malware anymore, it's prompts.

— George Kurtz, No Priors
Key Insight
Signature-based defense is a lookup table of known-bad; prompt-generated attacks defeat it by construction. That is why Kurtz's bet is a model that reasons about behavior, not a database that matches files.
03The New Speed

Breakout Time Is Now 27 Seconds

Breakout time, the window from landing on a machine to pivoting toward the crown jewels, has collapsed from days to as little as 27 seconds — the speed of inference.

On average, it's 29 minutes, but we've seen it as quick as 27 seconds.

— George Kurtz, No Priors
Key Insight
The jump from days to seconds means attacks now move at what Kurtz calls the speed of inference, which is the case he makes for defense that runs just as fast rather than at human pace.
04The Supply Side

No Guardrails at Home

Frontier guardrails are so strict they block a cybersecurity PowerPoint, so attackers skip them entirely and run ablated open-weight models on hardware at home.

You want to create ransomware? No problem. You want to create a ransomware kit? No issue.

— George Kurtz, No Priors
Key Insight
There is an asymmetry here: the same guardrails that refuse a defender's request do not stop an attacker, who simply drops to an open-weight model. The restriction meant to contain misuse ends up constraining mainly the people trying to defend.
05The Defense

Give the Defender Frontier-Grade AI: Safe Mind

CrowdStrike's answer is Safe Mind: red and blue AI models that attack and defend each other inside a digital twin of your network until the defender wins.

My biggest takeaway was that the defenders didn't have the same capabilities as the offense.

— George Kurtz, No Priors
Key Insight
The red-and-blue loop is self-play pointed at security: the edge is less a smarter model than the chance to rehearse the attack on a twin of the customer's real environment, over and over, before it happens for real.
06The Moat

The Moat Is the Data, and It Costs 3 Cents

A frontier model charges about $10 per remediation; CrowdStrike's own model does it for 3 cents, because it is post-trained on 15 years of attack data no lab can scrape.

there is no Reddit of CrowdStrike security data.

— George Kurtz, No Priors
Key Insight
In security the moat quietly moved from model quality to proprietary telemetry: a lab can copy the architecture but not the data, so specialization plus owned telemetry beats raw frontier scale on both cost and fit.
07The Real Exposure

They're Not Breaking In — They're Logging In

Even with zero code vulnerabilities you still get breached, because attackers are not breaking in, they are logging in, and every agent is a new superhuman identity to abuse.

they're not necessarily breaking in, they're logging in. There's such an issue around identity.

— George Kurtz, No Priors
Key Insight
Agents collapse the identity perimeter — Kurtz describes each one as a superhuman identity with its own compute, data, and network access — so the defender's control plane shifts from patching code to governing machine identities.
08The Reframe

Stop the Doomerism: An Agent Doesn't Beam Itself Anywhere

The agent threat is more attacks, faster — not a new class of attack; an agent still does what a human attacker does, just better and faster.

an agent doesn't just beam itself from your computer to another computer. It actually just does the same thing a human would do, just better and faster.

— George Kurtz, No Priors
Key Insight
If the techniques are unchanged and only volume and speed scaled, then defenses that already work can win by scaling the same way — which is why Kurtz reframes security as the gas pedal for AI adoption, not the brake.