OpenAI published a safety and security update on 16 September 2024 detailing changes to its Preparedness Framework and the establishment of a new Safety & Security Committee [source].

The company announced it had completed its 90-day review of safety processes following the dissolution of its previous Superalignment team in May 2024. The update confirmed that OpenAI's Board now oversees safety and security matters through the newly formed committee, which includes board members and technical advisors.

Key changes include updated catastrophic risk thresholds in the Preparedness Framework. OpenAI stated it would not deploy models assessed at "high" risk for cybersecurity, chemical/biological/radiological/nuclear threats, persuasion, or model autonomy without additional safeguards. The framework now includes a "critical" risk tier above "high" for future assessments.

The company reported that GPT-4o was assessed as "medium" risk across all categories under the revised framework. OpenAI stated it conducts red-teaming exercises and external evaluations before major releases, though specific evaluation methodologies were not detailed in the announcement.

The update followed public scrutiny of OpenAI's safety practices after the departure of several researchers earlier in 2024. The company stated it had expanded its safety systems team and established processes for escalating concerns to leadership and the board.

OpenAI indicated it would publish updated Preparedness Framework documentation and continue quarterly transparency reports. The announcement did not address whether the new oversight structure would apply retroactively to models already in production or specify enforcement mechanisms for the stated risk thresholds.