Anthropic open-sources Bloom and Petri for behavioral auditing of frontier AI models

Anthropic has released two open-source tools for probing how frontier AI models behave in practice: Bloom, a framework for automated behavioral evaluations, and Petri, a companion auditing framework for examining how risks combine. According to a dev.to write-up, the release ships with the underlying code, seed configurations and a technical account of how evaluation suites are generated, executed and judged — turning a previously internal methodology into something outsiders can run and adapt.

How Bloom builds an evaluation

The Bloom technical report, as summarized on dev.to, lays out a four-stage pipeline named Understanding, Ideation, Rollout and Judgment. The workflow starts from a seed configuration, develops candidate evaluation ideas, executes them against target models, and then judges the resulting behavior. The point is repeatability: a meaningful test needs a defined behavior of interest, scenarios that can be rerun, and criteria for interpreting what comes back.

Anthropic reportedly applied the pipeline across 16 frontier models and four behaviors: delusional sycophancy, instructed long-horizon sabotage, self-preservation and self-preferential bias. The report also documents metrics such as elicitation rate, which tracks whether a suite actually surfaces the behavior it targets, and suite diversity, a measure of variation within the generated tests. These describe properties of an evaluation suite, not a general safety ranking for any model.

Petri looks at risk interactions

Petri fills a different gap. Where Bloom measures whether a specific behavior appears, Petri is built for auditing how risks interact, exploring many scenarios in parallel. That distinction matters in deployment: a failure often emerges from a combination of conditions — a particular prompt plus particular tool access plus a particular workflow — rather than from any single input in isolation.

Licensing, installation and stewardship

The Bloom codebase is published under the MIT license, and the repository includes installation guidance covering pip installs from GitHub, the use of seeds and configurations, and the wiring needed to connect target models. The dev.to piece dates Anthropic's official Bloom announcement to December 19, 2025, and adds a stewardship caveat: as of 2026 the repository states it has a new home and is now developed and maintained by Meridian Labs. Teams considering it as a long-term dependency should review the current documentation first.

Part of a wider transparency push

A second dev.to article frames the release against Anthropic's broader disclosure record. It points to the company's Transparency Hub, model reports and system cards — notably the Claude Opus 4.5 System Card, which the article describes as the most direct first-party documentation of automated AI R&D evaluations — as evidence of a continuing measurement program rather than a one-off launch. The same article cites an arXiv preprint, "Measuring AI R&D Automation," that proposes three leading indicators: AI performance on research-like evaluations, how deeply AI tools influence high-stakes R&D decisions, and operational tracking of subversion incidents. Notably, that three-metric framework comes from corroborating academic research rather than Anthropic's own documentation, a distinction the article is careful to preserve.

What the tools do not promise

Neither tool is a safety certificate. An evaluation covers only the behavior, configuration, target model and judging approach it actually tests, so results do not certify a production application as safe. Teams deploying AI still need to validate their own prompts, tools, data access, escalation paths and customer-facing outputs. Meaningful runs also require access to the models being tested, infrastructure to execute the suites, and seed configurations that reflect risks relevant to the deployment.

Why it matters

Behavioral auditing of frontier models has largely lived inside labs, described in system cards and blog posts but difficult for outsiders to reproduce. By publishing Bloom and Petri with code, seeds and methodology under a permissive license, Anthropic makes that scrutiny practical: researchers can inspect how a behavioral benchmark was built instead of treating it as a black box, and AI teams can adapt the pipeline to risks specific to their workflows. The release also gives buyers a concrete vocabulary for interrogating vendors — how a behavior was tested, how varied the tests were, how results were judged — which is more actionable than a blanket assurance that a model has been evaluated. The open question is stewardship: with Bloom now maintained outside Anthropic, the framework's long-term direction depends on continued investment from its new maintainers.