The nonprofit watchdog Midas Project claims that OpenAI violated California’s SB 53 law at least three times during 2026. The allegations, detailed in an analysis dated September 10, 2026, focus on the failure to publish mandatory loss-of-control risk assessments for the GPT-5.6 and GPT-6 Astra models.
California’s Transparency in Frontier Artificial Intelligence Act, or SB 53, took effect on January 1, 2026. It requires developers to publish safety frameworks addressing catastrophic risks, including autonomous AI behavior. OpenAI’s own Frontier Governance Framework mandates that system cards for covered models must include a specific loss-of-control risk tier.
The Midas Project reports that this tier is missing from the system cards for GPT-5.6 and GPT-6 Astra. Additionally, the watchdog cited an incident from February 2026 involving the GPT-5.3-Codex model, which allegedly lacked necessary misalignment safeguards.
Released on September 3, 2026, the GPT-6 Astra model is classified under the Critical cybersecurity capability tier, the highest level in OpenAI’s internal framework. The watchdog argues that such a classification necessitates a comprehensive loss-of-control assessment.
OpenAI has denied these allegations, maintaining that it strictly complies with SB 53 and has confidence in its internal testing and governance processes. If the claims are substantiated, the company could face civil penalties under the state law.
