Artificial Intelligence & Machine Learning
,
Black Hat
,
Events
NIST’s Apostol Vassilev Explains Need for Dynamic Response, Not Static Testing
As artificial intelligence models grow in scale and power, leading to even more unpredictable outcomes, security teams are grappling with how to defend technologies that some experts can’t begin to fully comprehend.
See Also: AI Agents Demand Scalable Identity Security Frameworks
Static testing isn’t going to catch all potential vulnerabilities, which is why cyber response teams are exploring the practice of continuous red teaming, said Apostol Vassilev, research team supervisor at the National Institute of Standards and Technology.
“You have to apply proactively red teaming to change the state of your model such that attackers will have difficulty finding the latest adversarial prompts that will attack your specific instance,” Vassilev said. “Make the game dynamic so that you make the life of the attacker a little harder.”
In this video interview with Information Security Media Group at Black Hat USA 2025, Vassilev also discussed
- Information overload as an attack vector;
- The challenges of validating complex, natural language inputs;
- Combining traditional cyber practices with AI-specific measures.
Vassilev, a research team supervisor at NIST, focuses on adversarial machine learning and robust AI, Vassilev works closely with academia, industry and government agencies to develop and adopt standard for AI. Vassilev was previously awarded the bronze medal by the U.S. Commerce Department and has been profiled in the Wall Street Journal, Politico, Forbes and many other publications.