Overview
Severity: MEDIUM | Affected: US AI Safety Institute (AISI) | Category: policy
The U.S. AI Safety Institute (AISI), under the Department of Commerce, has officially released its 'Secure AI Framework and Evaluation' (SAFE-1) standards. A key provision of this new framework is the mandatory incident reporting requirement for developers of 'dual-use foundation models,' as defined by Executive Order 14110. Companies meeting the criteria must now report significant safety incidents, including critical model failures, large-scale jailbreaks, and evidence of models being used for malicious purposes, to AISI within 72 hours of discovery. The framework aims to create a centralized repository of safety-relevant data to help the government and the broader research community understand systemic risks and develop effective countermeasures. While praised by safety advocates, some industry groups have expressed concerns about the broad reporting criteria and potential disclosure of proprietary information. The first compliance reports are due by Q3 2025.