According to Anthropic (@AnthropicAI), their recent experiment with synthetic jailbreaks showed that Constitutional Classifiers significantly reduced the effectiveness of jailbreaks. The classifiers increased refusal rates by a small amount (+0.4%) while also raising compute overhead by 24%. These findings are essential for traders focusing on AI technology investments, as they highlight the potential efficiency and cost implications of deploying such classifiers. Anthropic is actively working on reducing these increased costs, which could impact future AI adoption and market dynamics.