Anthropic launches Bloom to help researchers understand how AI models behave in real situations

Anthropic

Anthropic has launched Bloom, a new open-source tool designed to help researchers better understand how advanced AI models behave in real-world situations, especially when things don’t go as planned.

As AI systems become more powerful and are used in increasingly complex settings, questions around alignment and safety have become harder to answer. Traditional evaluations often take weeks or months to build and can quickly become outdated. Models may learn the test itself, or their capabilities may improve so much that older evaluations no longer reveal meaningful behaviour. Bloom is Anthropic’s attempt to fix that problem.

Instead of relying on fixed test cases, Bloom automatically generates fresh evaluation scenarios for a specific behaviour defined by a researcher. The goal is to measure how often a model shows that behaviour, and how severe it is, across many different situations. This allows researchers to move faster while still getting results that reflect how models behave outside carefully controlled demos.

Bloom focuses on behaviour that matters for AI safety. With the launch, Anthropic has shared benchmark results for four alignment-relevant behaviours: delusional sycophancy, instructed long-horizon sabotage, self-preservation, and self-preferential bias. These tests were run across 16 frontier AI models, and according to Anthropic, the results closely matched what human evaluators would conclude.

The tool works through a four-step automated process. First, Bloom analyses the behaviour the researcher wants to study and builds an understanding of what should count as that behaviour. It then generates multiple scenarios designed to trigger it. These scenarios are played out through simulated conversations, after which a separate judge model scores how strongly the behaviour appeared. Finally, Bloom produces overall metrics such as how often the behaviour was elicited.

One key advantage of Bloom is flexibility. Each evaluation run creates new scenarios, which reduces the risk of models overfitting to a known test set. At the same time, results remain reproducible through a shared configuration file, allowing researchers to compare findings reliably.

Anthropic says Bloom complements an earlier open-source tool called Petri, which explores AI behaviour across broader multi-turn conversations. While Petri looks for many potential issues within a scenario, Bloom narrows in on one behaviour at a time and measures it in depth.

To validate the system, Anthropic tested Bloom on AI models that were intentionally designed to show odd or misaligned behaviour. In most cases, Bloom successfully distinguished these models from standard ones. The company also compared Bloom’s scores with human judgments and found strong agreement, particularly for cases where behaviour was clearly present or absent.

Bloom is now available publicly on GitHub, and early users are already applying it to study jailbreak vulnerabilities, evaluation awareness, and other safety concerns. Anthropic says tools like Bloom are becoming essential as AI systems move from labs into real-world environments, where understanding how models actually behave matters just as much as what they can do.