Reflection AI publishes Beam safety results and its Open Safety Framework
Article timeline
The articles in this story. Times are ET.
ReflectionOfficialReflection AI describes its Open Safety Framework for safe releasesAIReflection AI says it developed the Open Safety Framework, a methodology covering systemic risks, refusal behavior, and adversarial robustness. The company says the framework is meant to ensure safe and responsible releases at scale.
- ReflectionOfficialReflection introduces Iris Safety Evaluations to test Beam's safety policy adherence
AIReflection developed the Iris Safety Evaluations (ISE) to test whether Beam's responses align with its safety policy in adversarial and agentic settings. Across cyber offense, chemical, biological, and radiological/nuclear categories, Beam has higher refusal rates than leading Western and Chinese open models, according to the company.
- ReflectionOfficialReflection AI's Beam declines 89% of harmful requests in safety tests
AIReflection AI reports that Beam declines 89% of harmful requests and responds to 82% of safety-related harmless requests. The company says this is the highest refusal rate among the Western and Chinese open-weight models it tested.
- ReflectionOfficialReflection AI explains how it trained and evaluated Beam for safety
AIReflection AI says it has published a blog post on how it trained and evaluated Beam, its first model, for safety. The company says it will release additional safety research, evaluations, and tools under Project Iris, its open AI safety initiative, in the coming weeks.