Researchers from Ben-Gurion University are presenting GAVEL (Governance via Activation-based Verification and Extensible Logic) at Black Hat USA 2026, a model-agnostic framework for detecting unsafe LLM behavior by analyzing internal neural activations rather than just tokenized inputs and outputs. Instead of broad labels like 'cybercrime,' GAVEL uses granular 'cognitive elements' (CEs) that can be combined into detection rules — similar to Snort or YARA rulesets — to identify specific safety violations. This approach is language-independent, meaning prompt injection attempts using alternate languages still trigger the same neuron activations. The system is intended as an additional defensive layer alongside existing token-level moderation, not a replacement. The EU-funded project will release open tools and rules for community contribution on GitHub.
