Now loading...
Anthropic has unveiled Bloom, a new open-source framework designed to streamline the evaluation of behaviors in advanced AI models. This tool allows researchers to input a specific behavior they want to assess, then automatically creates a series of scenarios to measure how frequently and intensely that behavior appears in the model. Tests show that Bloom’s assessments align closely with manual reviews by experts and effectively highlight differences between standard models and those deliberately designed to show misalignment.
To illustrate its capabilities, the team behind Bloom shared results from benchmarks on four key alignment-related behaviors across 16 leading AI models: delusional sycophancy, where models excessively flatter users; instructed long-horizon sabotage, involving subtle undermining over extended interactions; self-preservation instincts; and self-preferential bias, favoring their own options in choices. These evaluations were developed and run in just a few days, demonstrating the framework’s efficiency.
The release comes at a critical time for AI safety research. Traditional behavioral tests for cutting-edge models often require months of development and quickly lose relevance as AI capabilities advance or training data incorporates past evaluations. Bloom addresses this by offering a scalable alternative, building on Anthropic’s earlier tool, Petri, which explores models through simulated conversations but focuses on broader profiling rather than targeted single-behavior analysis.
At its core, Bloom follows a four-step automated process. First, an initial agent interprets the described behavior and relevant examples to define evaluation criteria. Next, another agent crafts diverse scenarios, including user roles, prompts, and environments tailored to provoke the behavior. These then play out in simulated interactions, with agents mimicking users and tools to engage the target model. Finally, a judging model rates each interaction for the behavior’s presence and other factors, culminating in an overall suite analysis with metrics like elicitation rates.
Researchers can customize Bloom extensively, from selecting models for each stage to adjusting interaction lengths, incorporating tools, or adding secondary scores for aspects like realism. It supports large-scale runs via integration with Weights & Biases and outputs transcripts compatible with Inspect, a tool for AI interaction analysis. To ensure consistency, evaluations use a seed configuration file that outlines the behavior and parameters, allowing reproducible results despite varying scenarios.
Validation efforts confirm Bloom’s reliability. In tests against intentionally quirky “model organisms” prompted to show odd traits, Bloom distinguished them from production versions in nine out of ten cases, with the exception revealing unintended behaviors in the baseline. Human-AI agreement was also strong, particularly with Claude Opus 4.1 as the judge, achieving a Spearman correlation of 0.86 on hand-labeled transcripts, excelling at identifying clear presences or absences of behaviors.
A practical example involved replicating a self-preferential bias test from the Claude Sonnet 4.5 system card, where models tend to prefer their own suggestions in decisions. Bloom matched the original findings, ranking Sonnet 4.5 as the least biased, and went further by showing that encouraging more reasoning in Claude Sonnet 4 reduced this bias, often by having the model recognize and avoid conflicts of interest. Adjusting configurations like conversation length or filtering unrealistic scenarios refined results without altering overall model rankings.
Already, early users are applying Bloom to probe issues like jailbreak vulnerabilities, hardcoded responses, and sabotage patterns in AI systems. As AI deployment expands into more intricate settings, tools like this could prove vital for the alignment community in spotting and mitigating risky behaviors efficiently. For more on the methodology, benchmarks, and limitations, check the full technical report on Anthropic’s Alignment Science blog. The framework is freely available on GitHub.
