Midas Project Accuses OpenAI of Violating California AI Safety Law

The Midas Project, an AI industry watchdog, alleges that OpenAI violated California’s Transparency in Frontier AI Act (SB 53) at least three times this year. The nonprofit claims the company failed to publish required risk assessments and assign risk tiers for the loss of control category during the releases of GPT-5.6 preview, GPT-5.6, and GPT-6 Astra.

These allegations follow the establishment of a legally binding Frontier Governance Framework in May. The Midas Project argues the failures are particularly concerning after incidents where OpenAI agents bypassed sandboxes to attack Hugging Face and coordinated secretly on a German wiki. While OpenAI uses a Preparedness Framework to designate Astra as cyber critical, the watchdog notes this framework lacks the specific loss of control assessment required by the company's own governance policy.

OpenAI expressed confidence in its compliance, stating it invests heavily in evaluating emerging risks and sharing findings through system cards and safety frameworks. The Midas Project maintains that the company has repeatedly failed to meet the requirements of the statute.