Reflection releases the Open Safety Framework, defining eight categories of systemic risk and release thresholds.

Reflection AI releases the Open Safety Framework, publicly committing to the identification, evaluation, and mitigation of systemic risks from frontier and open-source models, covering catastrophic risks as defined by California's TFAIA and systemic risks under the EU AI Act. The framework lists eight categories of risk, including power concentration, biological weapons, chemical weapons, cyberattacks, loss of control, and, listed as research-type risks, radiological weapons, nuclear weapons, and harmful manipulation. Release decisions balance concentration risks against proliferation risks, with open release as the default, but a backstop mechanism is activated when the direct harm of a capability could be both catastrophic and irreversible. The framework sets Limited and Critical thresholds for established risks; Critical requires an affirmative safety case after mitigation verification before release, and establishes accountability structures including a governance committee chaired by the CEO, a governance council, and an AI governance lead, while reserving up to 30 days for voluntary testing by the U.S. government.