The Uncomfortable Math of Frontier AI in Production

Your firewall looks for bad intent. The new attacks don't have any, just malicious geometry that forces a chosen output and looks like garbage to every detector you own. It transfers across models. Add agents that already hold wallets, transacting at millisecond speed with no human in the loop, and the attack surface becomes the economy itself.

The uncomfortable part: deterministic attacks versus probabilistic defenses is not a fair fight. Prevention isn't weak, it's outmatched. The question shifts from whether you can catch every threat to whether the business can recover when you don't.

Drawing on Rubrik Zero Labs research and global CXO engagements, Kavitha Mariappan examines what it takes to secure frontier AI in production. Prevention alone is no longer a strategy. Resilience is what keeps the business standing.