Watchdog: OpenAI broke SB 53 three times with GPT-6 Astra

California's SB 53 gave the state a narrow role: it does not dictate what AI safety rules companies must set, only that they follow the rules they publish. OpenAI complied on paper in May, releasing its Frontier Governance Framework with four risk categories and three tiers of mitigation. The Midas Project's complaint is about what came after. None of the models OpenAI shipped since — GPT-5.6 preview, GPT-5.6, and GPT-6 Astra — carry the tier scores that framework promised, according to the watchdog. OpenAI points instead to its separate Preparedness Framework, under which Astra is rated cyber "critical," meaning it can autonomously execute advanced cyberattacks. The Midas Project notes that this alternative framework has no loss-of-control assessment, even as rogue-agent incidents have drawn attention to exactly that risk. OpenAI itself has acknowledged the limits of the current law, asking California in August to add monitoring requirements for models during training and evaluation, not just after deployment. The outcome hinges on whether state regulators treat the gap as a violation, since the incidents the watchdog cites were not required to be reported under the law. Until New York's RAISE Act takes effect early next year, California stands alone in requiring frontier developers to adhere to their own safety commitments — so how this plays out may shape whether that model of regulation holds.