Frontier AI has reset the balance of power in security, and so far the advantage has run one way. Attackers can pick up general-purpose models at will, even though no one built them for defenders, pushing the industry toward purpose-built systems trained on security data rather than the open web.
That shift is now taking shape as a partnership between CrowdStrike Holdings Inc. and Nvidia Corp, with autonomous red teaming making its debut alongside a defensive counterpart and the harness that runs them. The result is a model family aimed squarely at security operations, according to Daniel Bernard (pictured, right), chief business officer of CrowdStrike.
“Frontier models have done a fantastic job bringing AI innovation to the market at large. It’s really benefited the adversary,” Bernard said. “It’s time for the defenders to have something, and it’s time for security to have its own model. It turned into a set of models, a model family, and that’s where SafeMind was born.”
Bernard and Justin Boitano (left), vice president and general manager of enterprise computing at Nvidia Corp., spoke with theCUBE’s Dave Vellante at Fal.Con, during an exclusive broadcast on theCUBE, SiliconANGLE Media’s livestreaming studio. (* Disclosure below.)
Frontier AI meets a purpose-built harness
The offensive and defensive models are post-trained on Nemotron, Nvidia’s open-weight family, then run against a digital twin of a customer environment. In evaluations against leading frontier and open-source models, SafeMind delivered a 29% higher detection rate, six times faster remediation and 99% lower cost, according to CrowdStrike’s launch announcement. Nvidia has been a design partner and has applied the approach internally, Boitano noted.
“The digital twin describes the environment of the actual world,” Boitano said. “You run the red agent through the environment and you’ll find different ways in to exfiltrate data. Then the blue agent will come in and write rules that would have detected or prevented the red attack agent from getting through. That iterative loop basically hardens the environment.”
Packaging follows the same logic, with the models landing natively in the Falcon platform and the harness sold separately for teams automating security operations, Bernard explained. That arrives alongside Falcon IQ, which productizes the company’s partner-led assessment playbook for frontier AI risk.
“Our job in the market is [to] calm everybody down, and the way we’re going to calm everybody down is we’re going to stop more breaches than we’ve ever stopped before,” Bernard said. “There’s no reason to fear the adversary. There’s no reason to fear AI. You’ve got to make sure you have the right protection, the right systems, the right people, the right processes.”
Here’s the complete video interview, part of SiliconANGLE’s and theCUBE’s coverage of the Fal.Con event:
(* Disclosure: TheCUBE is a paid media partner for the Fal.Con event. Neither CrowdStrike Holdings Inc., the sponsor of theCUBE’s event coverage, nor other sponsors have editorial control over content on theCUBE or SiliconANGLE.)