Frontier Model Forum

Managing and understanding these risks effectively requires appropriate information flows across the ecosystem, ensuring timely notification of actionable information to the right sets of actors.

Timelines for incident reports are often set with well-intentioned urgency, but if they implicitly assume that root-cause analysis or full impact assessment has already been conducted, they will either produce premature and unreliable reports or plac...

Develop and publish monitorability evaluations. Standardized evaluations and metrics can improve assessments for how monitorable an AI model’s reasoning process is. These tests would measure the clarity, coherence, and faithfulness of a model’s Chain...

[...] if the evaluation results indicate [...] an exploitable vulnerability that may lead to a significant increase in biorisks, this information should not be published.