OpenAI wrote California's AI safety law. Then it ignored its own rules. At least 3 violations this year.

OpenAI published a legally binding safety framework under California's SB 53 in May.

It requires a loss-of-control risk tier for every model release.

Three models later, none of them have one.

GPT-5.6 Preview. GPT-5.6. GPT-6 Astra. All shipped without the required loss-of-control assessment. All system cards reference a different, non-binding internal framework instead.

Astra was classified as "Critical" for cybersecurity. The highest tier. The model can find zero-days and write exploits autonomously. But OpenAI never documented whether it can be reliably shut down.

The Midas Project alleges at least 3 violations this year. Penalty: up to $1M per violation.

Here's the structural problem. SB 53 is the only U.S. frontier AI law with enforcement teeth. It relies entirely on self-assessment. The company writes the rules, then reports whether it followed them. OpenAI wrote the rules. Then didn't follow them.

The same week OpenAI's chief scientist called for "widely mandated safety bars," the mandated bar already existed. OpenAI just wasn't using it.

And OpenAI is lobbying California to strengthen the law. While failing to comply with the current one.

If the world's most prominent AI company can't follow the lightest-touch regulation on the books, your vendor risk framework just became your most important document.

Audit every AI provider's compliance posture against SB 53 today. If they can't show you their risk tier assessments, you're buying unregulated critical infrastructure.

SOURCE: https://fortune.com/2026/09/14/openai-may-have-violated-californias-ai-safety-law-with-astra-model-releases-ai-watchdog-says/
VERIFIED: Fortune, Model Republic/Midas Project, Cloud Security Alliance
SIGNAL: This is the first real enforcement test of the only U.S. frontier AI law with penalties. If OpenAI can't comply, no vendor can. Every enterprise deploying frontier models is now exposed to regulatory risk they can't mitigate.