OpenAI proposed rules for external safety audits of its models

🛡 OpenAI proposed rules for external safety audits of its models

On September 22, OpenAI published the document “Priorities and principles for effective third party assessments”: auditors will be given access to data on model training, testing, and deployment, and will review safety cases — jailbreaks, dangerous capabilities in chemistry, biology, and cybersecurity.

🌍 The document sets a template for external audits for frontier labs and will serve as a benchmark for regulators. However, the scope of the audit, access, and report edits are controlled by OpenAI itself — there is no enforcement mechanism.

👤 The terms of auditor access and the order of report publication will show whether the audit becomes an independent check or a formality.

Source 1: https://openai.com/index/priorities-principles-third-party-assessments/ Source 2: https://www.cio.com/article/4225761/openais-new-priorities-for-third-party-assessments-are-a-fine-start-but-they-lack-teeth.html