OpenAI Urges Stronger AI Regulations in California, Marking a Shift in Stance

Calls for Expanded SB 53 Safety Regulations
Shift from Opposition to State-Level Rules
Emphasis on Strengthening Safeguards After AI Model Security Incidents

OpenAI, the developer of ChatGPT, which has previously expressed a negative stance toward state-level artificial intelligence (AI) regulations, has now called for stronger AI safety regulations in California. This shift is being attributed to the fact that cutting-edge AI cyber threats are becoming a reality, including a recent incident in which OpenAI’s own model escaped a controlled test environment and accessed an external system without authorization. As a result, OpenAI’s approach to regulation appears to be changing.

OpenAI LinkedIn Post. LinkedIn

According to LinkedIn, OpenAI’s Government Affairs team posted on August 21 (local time) that California’s AI regulatory bill, SB 53, “should be amended to expand safety measures.”

OpenAI added that the series of recent incidents “demonstrates the need for such safeguards and the importance of updating them in response to emerging risks and new safety challenges.”

SB 53 is the Transparency in Frontier Artificial Intelligence Act, signed into law by California Governor Gavin Newsom on September 29 of last year. The act requires major advanced AI developers to publicly disclose their own safety management frameworks and report significant safety incidents to the state government, while also protecting whistleblowers who expose AI-related risks.

The current law defines large-scale risks posed by cutting-edge AI—including AI-enabled hacking, biological attacks, and loss of developer control—as subject to management and regulation. Major AI developers must make public how they assess and mitigate model risks, and explain how they protect unreleased model weights from cyberattacks in their safety frameworks.

OpenAI has taken its demands a step further, arguing that the scope of safety regulations should be expanded. Specifically, OpenAI proposed that it should be mandatory to monitor whether advanced AI models that are still in the training or evaluation stages could bypass third-party security controls or compromise confidential information, thereby potentially causing major incidents. The company also argued that cybersecurity measures should be strengthened across the entire development process, from model training to evaluation and deployment.

Through these measures, OpenAI explained, developers would be able to detect risks early and respond quickly, in addition to sharing best practices that could enhance safety across the industry as a whole.

One Year Ago: Opposed to State Regulation, Now Changing Position Amid Rising AI Cyber Threats

OpenAI’s current position stands in stark contrast to its stance just one year ago, when it opposed state-level AI regulation. Chris Lehane, OpenAI’s Chief Government Affairs Officer (CGAO), warned Governor Newsom in an open letter last August that if the United States failed to establish unified federal AI guidelines and each state adopted different rules, innovation could be hindered.

This change in OpenAI’s position appears to be driven by mounting concerns over the rapidly increasing cyberattack capabilities of AI models. Last month, OpenAI revealed that during internal security tests, one of its AI agents escaped a quarantined test environment and accessed the system of the external AI platform Hugging Face without authorization. After this, OpenAI reinforced isolation protocols and real-time monitoring for high-risk AI tasks, and even temporarily paused reinforcement learning training on some cutting-edge models.

Nevertheless, OpenAI’s recent stance does not necessarily represent a complete pivot to state government regulation. On July 15, OpenAI announced a policy proposal introducing the concept of “reverse federalism.”

This approach suggests that, while the adoption of comprehensive federal AI regulations remains delayed, individual states such as California and New York should first establish core safety standards that are mutually compatible, paving the way for the development of a unified federal regulatory framework in the future.

This content was produced with the assistance of AI translation services.