George Kurtz, CrowdStrike's CEO, says AI agents have flattened the threat pyramid, letting hacktivists wield capabilities once reserved for nation states.
CrowdStrike's breakout time has fallen to minutes, averaging 29 minutes with the fastest case at 27 seconds, Kurtz says.
Kurtz says defenders were left behind in the Hugging Face incident because frontier models refused to help, forcing CrowdStrike to rely on open weight models for its Safemind defense.
CrowdStrike's CEO says AI agents made attackers faster, and defenders must match
AI agents have given every attacker the speed and skill once reserved for nation states, and breakout times now run in minutes. Frontier models refuse security work, so adversaries use open weight models while defenders get blocked. CrowdStrike's answer is its own red and blue models, trained on years of proprietary security data, which Kurtz says run far cheaper than frontier models. In his view defenders keep a mathematical edge if they build for the agent era.
Takeaways
AI has flattened the adversary pyramid. Hacktivists sit at the base of Kurtz's pyramid and nation states at the top. Models without guardrails give hacktivists nation-state capabilities. Attacks arrive faster in a constant stream, and breakout time now averages 29 minutes, with the fastest seen at 27 seconds.
Attackers use open weight models, defenders get refused. Frontier models refuse cybersecurity requests, so adversaries lean on open weight models. After the Hugging Face incident, which Kurtz calls the first agentic end-to-end attack, frontier models refused to help analyze it and the team turned to open weight ones.
CrowdStrike built red and blue models with NVIDIA. Safemind pairs Red Tempest and Blue Solano, built over about six months. A digital twin of the customer's environment lets the red model keep attacking until the blue model learns its moves. Nemotron is a main model on the blue side.
Lower cost and security data are the edge. CrowdStrike's model costs about $0.3 for general remediation, against about $10 for the top frontier cyber model. Kurtz credits about fifteen years of security data, including managed detection, and roughly 7tn events a day from its sensors.
Agents must be discovered and fully logged. CrowdStrike's platform has grown from one module to 33, covering endpoint, cloud, IT and identity. Its first aim with agents is finding rogue and shadow AI, then logging an agent's full chain of events, from sub-agents to MCP server access.
Air gapping critical systems is no fix. Kurtz calls the idea nonsense. Agents do what humans do, only faster, and AI has mainly created more vulnerabilities. Even zero vulnerabilities would not stop breaches because identity is a main way in, and legacy software such as Microsoft's leaves a long tail.
Customers want sovereignty and good-enough models. Customers do not want their data used to train labs' models, though CrowdStrike would still route to frontier models if asked. A large bank's CIO found open weight models three months behind the frontier good enough. On distillation, Kurtz says the genie is out of the bottle.
What Kurtz expects next
Tap a name to go there · tap a line to see why they are linked
On this episode 3
Show 1
Local models will run on people's computers within two years, and open weight models will shrink enough to run on phones.
Incumbents should adopt smaller squads, change their operating cadence and protect moats such as a system of record.
AI will create jobs, and technology will appear in five years that nobody has yet thought of.
Cybersecurity is not a solvable problem, so students should study it.