Ongoing Safety Audits for AI Agents
On September 25, 2026, OpenAI CEO Sam Altman announced that the company is conducting an extensive and ongoing internal review regarding AI agents accessing the internet during training and evaluation phases. The statement, shared on X, aimed to clarify OpenAI's internal safety vetting protocols for AI systems capable of interacting with external networks.
Balancing Transparency and Verification
According to the post by Sam Altman (@sama), OpenAI has begun publishing periodic summaries of this audit process and pledged to share further updates. Notably, Altman acknowledged that the pace of public disclosures has 'not been as fast as we would like,' explaining that the company is striving to balance reporting speed with rigorous technical verification.
Technical Risks and Unanswered Questions
Granting AI agents direct internet connectivity and data retrieval capabilities during training or evaluation presents complex technical risks, including potential data leakage, training data poisoning, and unauthorized autonomous behaviors. However, OpenAI's brief announcement did not disclose quantitative metrics, details regarding affected model architectures, or specific incidents that prompted the wide-ranging review.
Altman's statement leaves several key technical questions unanswered. The announcement did not specify a timeline for completing the audit, the exact network access control mechanisms planned for future agent systems, or the concrete safety benchmarks required prior to production deployment.