Shifting AI Security Focus to Model Input-Output Controls

Check Point's Paul Barbosa on Implementing Runtime Defenses for AI Data Security Michael Novinson (MichaelNovinson) • December 10, 2025

Modern artificial intelligence systems execute actions through natural language instead of code, which creates new risks that traditional security controls cannot contain. Protecting AI at the point where users interact with models is essential, said Paul Barbosa, vice president for cloud security at Check Point Software.

See Also: Going Beyond the Copilot Pilot - A CISO's Perspective

Barbosa said a runtime layer offers the most stable point to enforce controls as underlying model architectures change rapidly. He tied the approach to prompt-injection persistence and to evasive tactics that bypassed static guardrails, including multilingual manipulation.

"Intersecting the input and output to the models is the appropriate place right now to look at where we can get the most effective security models for AI," he said. "We need to think about security in ...