Bloom automated behavioral evaluations
Bloom is an open-source agent framework that turns a named behavior into a suite of generated scenarios and measures its frequency and severity.
Complements Petri. Instead of open-ended auditing, a researcher specifies one behavior and Bloom generates scenarios, runs them and scores them. Anthropic released results for four behaviors on 16 models; its scores correlated strongly with hand labels and separated misaligned model organisms from baselines.
- Date
- Friday, 19 December 2025
- Lab
- Anthropic
- Kind
- infra
- Access
- paper only
The four behaviors are delusional sycophancy, instructed long-horizon sabotage, self-preservation and self-preferential bias.
Sources
This record was checked against its sources on 6 October 2026. How we check