OpenAI Faces SB 53 Compliance Allegations Over Missing Model Risk Tiers

OpenAI is facing a new challenge under California’s frontier-AI law after the Midas Project alleged that the company released GPT-5.6 models and GPT-6 Astra without publishing risk tiers required by its own safety framework. The potential stakes are significant: SB 53 can carry penalties of up to $1 million per violation, while OpenAI says it is confident it complies.

The allegation is not that OpenAI lacks safety evaluations altogether. It is that the company has published one set of evaluations while omitting a different set of disclosures it committed to make in its Frontier Governance Framework. California’s law requires major frontier-AI developers to publish a safety framework and adhere to their stated policies.

Two frameworks, one compliance question

OpenAI published the Frontier Governance Framework in May to explain how its safety and security practices align with California’s law. The framework covers four risk areas and says new models will receive tiers from one to three, with corresponding mitigation commitments. Fortune reported that the system cards for GPT-5.6 preview, GPT-5.6, and Astra do not publish those categories or tiers.

The framework’s four stated risk areas

• Cyber offense
• Chemical, biological, radiological, and nuclear risks
• Harmful manipulation
• Loss of control

OpenAI’s answer rests on the Preparedness Framework, which it describes as the basis for managing the most serious risks from advanced AI. Under that separate rubric, OpenAI designated Astra as cyber “critical,” its highest cyber-risk threshold. The company says the Frontier Governance Framework explains how those underlying practices align with regulatory requirements.

The missing category is the sharpest point

The two approaches overlap on cyber risk but do not map cleanly across every promised category. The Midas Project says OpenAI’s Preparedness Framework does not assess loss of control, even though that is one of the four areas in the governance framework. That makes the dispute more than a formatting complaint: a published cyber designation does not reveal whether OpenAI made a formal tier determination for loss of control.

Astra’s system card does address monitorability and controllability—whether a model’s actions can be observed and directed—and reports evaluations in those areas. But those discussions do not themselves publish the Frontier Governance Framework’s tier labels. The available documents therefore show safety work and safeguards, while leaving the claimed tier disclosure unresolved.

A young law meets its first hard boundary

Governor Gavin Newsom signed SB 53 in September 2025. It was designed to pair public safety frameworks with accountability for failing to follow them, rather than prescribing one universal technical test. That design leaves a practical enforcement question now: when a company uses a different internal framework, can it satisfy a public framework’s commitments without publishing the exact assessments that framework describes?

The Midas Project says OpenAI has violated the law at least three times this year. OpenAI disputes that conclusion. Until an enforcement authority addresses the allegation, the central fact remains contested—but the case exposes how much SB 53 depends on regulators deciding whether a company’s public framework is a binding checklist or a higher-level map to its broader safety program.