Palo Alto Networks advocates small AI models and machine-speed defenses against rising threats

Palo Alto Networks CEO Nikesh Arora says the growing capabilities of frontier, open-source and Chinese AI models are shifting corporate cybersecurity priorities. In his view, organizations will need to “fight AI with AI,” using systems that detect and respond to threats at machine speed. He estimates that modernizing cybersecurity architectures will take three to five years if companies commit to it.

Arora says Palo Alto Networks inspects about 180 TB of data daily, making it economically impractical to apply frontier models to every transaction. The company is therefore training small language models for specific tasks, such as classifying malicious websites, which can run on endpoints or laptops; he puts the training cost of one model at roughly $5,000–$20,000. More powerful defender models would be reserved for cases where a problem is detected.

He also says better AI-generated grammar is making phishing and social engineering more effective. Arora favors embedding small AI classifiers in email-scanning systems to assess sender addresses, domains and message context rather than relying solely on users. On NeoCloud security, he says he is not aware of the specific concern referenced, though he sees limited excess capacity and restricted connections to frontier-model customers as factors that may reduce the risk of compute hijacking; he adds that basic infrastructure security is still needed.