GAVEL: New AI Security Analyzes Model Internals

Researchers are developing a new way to secure AI systems. This method looks inside large language models, examining their activation patterns, rather than just their inputs and outputs. Offensive-security researchers will present this model-agnostic technique at Black Hat USA 2026. It uses standardized rules to process activation events, breaking down threats into specific "cognitive elements" like "create content" or "personal information." These elements can then be combined to detect specific malicious activities, such as phishing attacks. This approach, called GAVEL, aims to create an open system of identified cognitive elements and rules. What's interesting is that activation analysis is language-independent. This means it can detect malicious intent even if prompts are changed to bypass content filters. While still a research project, GAVEL could add a new layer of defense for AI systems.