Anthropic Releases Bloom, an Open-Source Framework for AI Behavior Evaluation

Anthropic has released Bloom, an open-source framework for automated behavioral evaluations of frontier AI models, according to an announcement on the company's website. The tool allows researchers to specify a target behavior and automatically generate scenarios that test how often and how severely that behavior appears.

Bloom operates through four stages—understanding, ideation, rollout, and judgment—to produce evaluation suites that quantify the presence of specific behaviors. It integrates with research tools such as Weights & Biases for large-scale experiments and exports results in Inspect-compatible formats. Each evaluation run can produce unique scenarios while maintaining reproducibility through a configuration seed file.

The company reported benchmark results for behaviors including delusional sycophancy, instructed long-horizon sabotage, self-preservation, and self-preferential bias across 16 models. Using Bloom, these evaluations were completed within days, and the results aligned closely with human-labeled judgments. Validation tests showed that Claude Opus 4.1 correlated most strongly with human scoring, achieving a Spearman correlation of 0.86.

Bloom complements Anthropic’s earlier open-source tool, Petri, which explores AI models' behavioral profiles through simulated interactions. Researchers can access Bloom and its documentation via the official GitHub repository.