OpenAI Asks Sacramento to Watch Models Still in the Eval Harness
OpenAI Global Affairs called for stronger SB 53 safeguards covering models in training or evaluation and cybersecurity across development. The policy post supports compatible state protections as a path toward a national standard. It is a proposal, not an enacted amendment.
OpenAI’s Global Affairs team called for California to expand SB 53 protections to cover serious incidents involving frontier models still in training or evaluation.
In its policy post, the team said monitoring should address conduct that could bypass a third party’s security controls and compromise confidential information. It also proposed stronger cybersecurity across the model-development lifecycle, including controls against models circumventing internal security measures.
What OpenAI asked for#
- Monitoring during training and evaluation, before deployment.
- Stronger lifecycle cybersecurity protections.
- Work with California’s legislature and governor on amendments.
- Compatible state safeguards that could contribute to a national standard, an approach the company calls “reverse federalism.”
TechCrunch’s August 22 report notes that OpenAI previously opposed SB 53, which established transparency and whistleblower protections for large AI companies.
Limits#
The inspected post refers generally to recent incidents. It does not name Hugging Face or admit that OpenAI’s Preparedness Framework failed. The policy position does not itself amend the law.
Correction (October 14, 2026): Linked and checked the original policy post; removed an unsupported attribution of a specific incident and admission of failure. The original publication date is retained; the supporting secondary report is dated August 22.