Emerging Practices in Frontier AI Safety Frameworks

Risk identification is moving beyond traditional testing: Companies are adopting "red-teaming" practices and pre-deployment evaluations to identify potential harms before systems are released, rather than relying solely on post-deployment monitoring. This represents a shift toward proactive rather than reactive safety approaches.

Governance structures are becoming more formalized: Emerging frameworks establish dedicated safety teams, cross-functional review processes, and clear escalation procedures for handling identified risks. This institutionalizes safety considerations rather than treating them as ad-hoc concerns.

Transparency and external accountability are key commitments: The frameworks emphasize publishing safety practices and engaging with external researchers and auditors, suggesting that companies recognize independent oversight as essential to building public trust in frontier AI systems.

Risk mitigation strategies are multi-layered: Rather than relying on a single safeguard, effective frameworks combine technical measures (like model testing), operational controls (like usage policies), and human oversight mechanisms to address different types of potential harms.

The field lacks standardized metrics: The paper highlights that safety frameworks are still developing, with no consensus yet on how to measure or compare the effectiveness of different safety approaches—suggesting this remains an active area requiring further innovation and collaboration.