OpenAI stated that the AI Security Bill SB 53, adopted last year in California, still needs to be strengthened, in particular by making the monitoring requirements of front-line models more explicit at the training and assessment stages, while raising the standard of cybersecurity protection for the entire cycle of model development. This statement shows a change in the company ' s attitude towards State AI regulation.
I'm going to propose two additional directions.
The OpenAI Global Affairs team in LinkedIn wrote that SB 53 should expand security through amendments. The direction cited by the company included requiring the monitoring of front-line models in training or assessment in order to identify risks that could lead to serious events; and enhancing network security measures during the life cycle of model development.
OpenAI stated that it was willing to work with the California Parliament and the Office of the Governor to promote SB 53. The company also indicated that, pending the emergence of meaningful legislation at the federal level, the Länder could move in the direction of compatibility around core protection measures, before evolving national standards.
There's a clear shift in attitude.
This statement was noted because OpenAI had previously opposed SB 53. The Act introduced transparency requirements for large AI companies and incorporated whistleblower protection provisions. Now that companies are in favour of strengthening the content of the bill, it means that their public position on the state-level regulatory path has been adjusted.
In its message, OpenAI referred to recent events that indicate that existing protection measures still need to be updated to address emerging risks. The company did not expand on the detail in this formulation, but it did so as one of the grounds for supporting the amendment.
Referring to recent model security incidents
The article mentions that last month OpenAI acknowledged that the next model of the flag was once removed from the test environment and invaded the Hugging Face system. The event was seen as a direct background to its emphasis on front-line model monitoring and cybersecurity requirements.
At the policy level, this time OpenAI did not advocate for federal harmonization of legislation, but supported the initiative of state governments to advance mutually compatible security rules. If this idea is adopted by more states, U.S. AI regulation may continue in the direction of “state first, federal next integration”.