OpenAI Safety and Security Practices Update

OpenAI has transitioned its Safety and Security Committee (SSC) into an independent Board oversight committee to govern critical safety and security measures during model development and deployment. This move establishes a formal mechanism for independent oversight of model launches, including the authority to delay releases until safety concerns are addressed.

Independent Governance and Board Oversight

The Safety and Security Committee will now function as an independent Board oversight committee. Chaired by Zico Kolter (Director of the Machine Learning Department at Carnegie Mellon University), the committee includes Adam D’Angelo (Quora CEO), retired US Army General Paul Nakasone, and Nicole Seligman (former EVP and General Counsel of Sony Corporation).

Key responsibilities of the committee include:

  • Model Launch Oversight: The committee and the full board exercise oversight over model launches. They have the authority to delay a release if safety concerns are not resolved.
  • Technical Briefings: Company leadership provides the committee with safety evaluations for major model releases. The committee receives regular reports on technical assessments for current and future models and post-release monitoring.
  • Operational Engagement: The committee maintains regular engagement with OpenAI’s internal safety and security teams.

Enhanced Cybersecurity Measures

OpenAI is adopting a risk-based approach to cybersecurity to protect advanced AI research and product infrastructure. Current initiatives include:

  • Infrastructure Security: Expanding internal information segmentation and increasing staffing for around-the-clock security operations teams.
  • Industry Collaboration: Evaluating the creation of an Information Sharing and Analysis Center (ISAC) for the AI industry to facilitate the sharing of threat intelligence and cybersecurity information among AI entities.

Transparency and External Collaboration

OpenAI is expanding its transparency efforts and third-party validation of its safety systems through the following channels:

Transparency Reporting

OpenAI continues to publish system cards for its models, such as those for GPT-4o and o1-preview, which detail capabilities, risks, external red teaming results, and frontier risk evaluations conducted under the Preparedness Framework.

External Partnerships

  • Independent Testing: The company is developing collaborations with non-governmental labs and third-party safety organizations for independent model safety assessments.
  • Governmental Cooperation: OpenAI is working with the U.S. and U.K. AI Safety Institutes to research emerging risks and standards for trustworthy AI. It is also collaborating with Los Alamos National Laboratory to study the safe use of AI in bioscientific research settings.

Unified Safety Framework for Development

OpenAI is implementing an integrated safety and security framework with clearly defined success criteria for model launches. This framework is based on risk assessments and must be approved by the Safety and Security Committee.

To support this framework, OpenAI has reorganized its research, safety, and policy teams to ensure tighter collaboration and stronger connections across the organization. This framework is designed to adapt as models increase in capability and complexity.

Sources