Anthropic releases Bloom, an open-source agentic framework that automates behavioral evaluations of frontier AI models. The tool generates targeted scenarios to quantify alignment risks like delusional sycophancy and self-preservation biases across models, achieving 86% correlation with human judgment.