Breaking Gemma 4 Safety AlignmentI extended the Silenced Biases (AAAI-26) research to Google's Gemma 4. Activation steering failed. Prompt-level attacks broke every bias category. Apr 8, 2026 AI Security, LLM Safety
