Why Does CrowdStrike Train Its Own Model?

George Kurtz says CrowdStrike trains its own security models for three separate reasons: the models are significantly cheaper than frontier models, post-training on CrowdStrike data focuses them on security, and customers want sensitive data kept off outside labs.

For general remediation, the highest caliber frontier cyber model probably costs about $10. CrowdStrike's is like three cents, a huge cost difference. The data advantage is separate from the price. There is no Reddit of CrowdStrike security data, and synthetic data is not the same as what CrowdStrike holds. He started 15 years ago, so the archive is 15 years or a little less, which he calls over a decade of security data, including the managed detection and response service, with threat and vulnerability pairings and attack chains. CrowdStrike trains on the massive set and is getting great results. He thinks the model is very focused because of the post-training. It is not going to cure cancer. It gives great results for security.

Customers have entrusted CrowdStrike with their data for the last 15 years. They do not want sensitive information going out to the labs or being trained on. AI sovereignty is very important to many of them. The labs have incredible technology, and CrowdStrike is not saying exclude them. A customer can use a frontier model, use CrowdStrike's models, or route between the models.