Startups Scramble to Build Immediate AI Security

This commentary surveys the state of AI/ML security, arguing foundational models have expansive attack surfaces and that traditional software security paradigms are insufficient; it highlights debates within MLSecOps (adversarial testing vs. securing the ML pipeline), references MITRE ATLAS, and profiles startups like HiddenLayer, Protect AI, Adversa AI, and Calypso AI. It also outlines the promise of fully homomorphic encryption for privacy-preserving AI while emphasizing its current performance and cost limitations, concluding that a small set of startups are shaping practical paths to AI security amid rising compute costs.