Strengthening the Safety Ecosystem Through External Testing

OpenAI actively leverages independent and trustworthy Third-party assessments to strengthen the safety ecosystem for frontier AI. The purpose is to verify claims about a model's safety capabilities and mitigation measures, prevent potential blind spots, and increase transparency.

External collaboration takes three main forms.

• Independent evaluations: Testing core risk areas such as biosecurity, cybersecurity, AI self-improvement, and scheming
• Methodology reviews: Verifying risk assessment and interpretation methods
• SME probing: Domain experts directly evaluate models through real-world tasks and provide feedback on safety measures

These evaluations serve as an independent layer that complements the limitations of internal testing and prevents Self-confirmation bias. For GPT-5, a wide range of external capability evaluations were coordinated to support deployment decisions, covering areas such as long-horizon autonomy, deception, oversight subversion, and potential for biological experiment planning.

The evaluation process utilizes benchmarks such as METR's time horizon evaluations and SecureBio's Virology Capabilities Test (VCT). OpenAI supports external organizations by providing secure access to early model checkpoints, a Zero-data retention environment, and models without safety mitigations applied, enabling precise measurement of a model's underlying capabilities.