AI Flattens Attack Barriers: Script Kiddies Now Wield Nation-State Capabilities
Open-source large models running on home computers can generate ransomware; after initial intrusion, attackers achieve lateral movement across networks in an average of 29 minutes, minimum 27 seconds.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Open-source models flatten the attacker pyramid: script kiddies gain nation-state capabilities
The traditional attacker hierarchy—with hacktivists at the bottom, cybercriminal organizations in the middle, and nation-state actors at the top—has been compressed by guardrail-free open-source models. Capabilities that were once nation-state exclusive are now available to any attacker. The result: more exploits discovered, faster attack velocities, and a tempo shift visible enough to distinguish AI-driven operations from manual ones.
— George KurtzFrontier models refuse security work; attackers deploy ablated open-source models
Frontier models now refuse to generate security-related content—even a PowerPoint on network defense will be rejected or downgraded to a weaker model. The models actually used at scale by attackers are open-source: GLM, Qwen, and others, many with safety guardrails intentionally removed or customized for red-team testing. Kurtz notes he keeps two Spark DGX machines at home with unrestricted access to all models—one step up from running everything on a laptop, with no barriers to building ransomware tools, packaging them, or iterating.
— George KurtzLateral movement from first foothold to network spread now averages 29 minutes
CrowdStrike tracks ‘breakout time’—the span from initial system compromise to lateral movement across the network. In the past, this was measured in days or hours. Now it averages 29 minutes, with a recorded minimum of 27 seconds. The speed comes from a shift in attacker method: instead of relying on pre-built malware, they use live prompt engineering in-battle. Each attack generates unique prompt-derived signatures that make fixed-pattern detection nearly impossible.
— George KurtzDefenders currently lack parity with attackers in frontier AI capabilities
When asked about the Hugging Face end-to-end agent attack research, Kurtz didn't highlight attacker survival tactics—he pointed to what the research team actually discovered: frontier models refused to help them. They had to turn to open-source models to complete their analysis. His conclusion: defenders don't currently have frontier-grade AI on the same level that attackers have built for themselves. This is precisely what CrowdStrike's self-developed models are meant to solve.
— George KurtzRed and blue teams fighting in digital twins prevents defender blind spots
CrowdStrike spent roughly six months in partnership with NVIDIA to build two proprietary models based on Nemotron: Red Tempest (red team) and Blue Solano (blue team). The method: encode a digital twin of each customer's IT environment, then let the red model continuously attempt breaches inside the simulation until the blue model detects and stops it. Jensen Huang insisted on the digital-twin layer—he argued that every other industry has them, so why not security? The continuous adversarial loop in simulation keeps defenses iterating faster than real threats.
— George KurtzAdvantage isn't cheaper inference; it's proprietary threat-to-vulnerability data pairs
A premium frontier security model costs approximately $10 per remediation task. CrowdStrike's model costs about $0.03. But cost is not the real differentiator. The moat is data: 10+ years of threat-to-vulnerability pairings and attack chains that public datasets cannot replicate. Kurtz frames it as ‘there is no Reddit without CrowdStrike security data’—the public internet lacks the density of real-world security incidents that make training effective. This is why enterprise customers refuse to send sensitive data to frontier-model providers for training; this is why ‘AI sovereignty’ matters.
— George KurtzToken costs force enterprises toward local post-training with open-source models
Kurtz recounts a dinner with a New York major bank's CIO: the company had bet everything on frontier models until the token bill became unsustainable. Now their strategy is to push the vast majority of work onto open-source models and post-train locally—even if they run three months behind the frontier, that delay is acceptable. Only high-return tasks—‘Formula 1 racing’ operations—justify the cost of frontier-model inference. Another large tech buyer was even more direct: the organization simply cannot absorb frontier models' iteration cycle (every 10 days). Running three months behind actually aligns better with their regulatory pace.
— George KurtzData creation from operational scale, not headcount, is the true moat
As AI-native startups prove that tiny teams can deliver what used to take thousands, large companies must self-disrupt or lose. Waterfall development—thousands of people, yearly releases—is no longer viable. The answer is small, moving squads. Kurtz himself joins three design reviews per week for Falcon Guardian and adjusts direction like a startup. The real moat isn't team size; it's being a ‘net data creator.’ The endpoint-security agent (not to be confused with an AI agent) generates approximately 7 trillion events daily, providing continuous training signal that neither startups nor frontier-model providers can match.
— George KurtzIn their own words · checked verbatim
For the first time, security is the gas pedal, not the brake.
George Kurtz0:00
the new apex predator is the agent state, not the nation state, it's the agent state.
George Kurtz3:11
On average, it's 29 minutes. But we've seen it as quick as 27 seconds.
George Kurtz7:17
it's not malware anymore. It's prompts. So they can basically engineer prompts that would uniquely create a totally unique signature of what direction they're going to go in, how they gather information, where they pivot.
George Kurtz8:25
there is no Reddit of CrowdStrike security data.
George Kurtz20:45
trust is, is built in drips and lost in buckets.
George Kurtz22:46
The only way to fight AI today is with better AI.
George Kurtz36:25
Figures
| Attacker lateral-movement time (breakout time) | Average 29 minutes, minimum 27 seconds | 7:17 |
| CrowdStrike platform module count | 33 | 16:44 |
| Average agents used per enterprise employee | ~90 | 17:44 |
| Cost per security remediation task | Frontier model ~$10; CrowdStrike model ~$0.03 | 20:45 |
| Daily events from CrowdStrike endpoint agent | ~7 trillion | 41:47 |
Glossary
- Pyramid of pain
- Attacker hierarchy ranked by sophistication and rarity: hacktivists at base, cybercriminal organizations in the middle, nation-state actors at the apex.
- Breakout time
- Time elapsed from initial system compromise to lateral movement across the target network.
- Ablated models
- Open-source models with safety guardrails intentionally removed to enable unrestricted content generation.
- Digital twin
- Virtual replica of a customer's IT environment used to simulate attack and defense scenarios.
- Distillation
- Training a smaller, cheaper model using outputs from a larger model.
How to listen
Enterprise CISOs, security team leads, and technology decision-makers who want to understand how AI is simultaneously reshaping both attack and defense capabilities and the cost structures that underlie them.
The ending on career advice—whether AI will displace workers, what fields kids should study—can be skipped; information density is low.