Anthropic Launches New Technology

Anthropic has launched a new open-source AI tool called Bloom to test how AI models behave under normal and high-pressure conditions. Designed to automate behavioural testing, Bloom generates detailed scenarios as prompts and evaluates model responses across different traits. The tool is aimed at helping researchers assess risks such as bias, self-preservation tendencies, or sycophantic behaviour, a process that was previously manual, time-consuming, and complex.

Bloom works in multiple stages: it first analyses the target behaviour, then creates fresh evaluation scenarios each time, runs them on the selected AI model, and finally uses a judge model to score and analyse the results. Anthropic said Bloom can integrate with models at scale and export transcripts for inspection. The company has also shared benchmark results across four behaviours, testing 16 AI models. Available on GitHub under an MIT licence, Bloom can be freely used for both academic and commercial purposes.