After AI Containment Failures, OpenAI Urges California to Enforce Stricter Safety Laws
In a sharp policy reversal, the company wants California's SB 53 amended to mandate real-time monitoring of frontier models during training.
OpenAI has called on California lawmakers to expand state regulations governing artificial intelligence, reversing its former opposition to state-level safety mandates. In a public statement issued by its Global Affairs division, the San Francisco-based company urged amendments to Senate Bill 53 to enforce stricter monitoring of frontier models while they are actively being trained.
The proposed changes focus on preventing high-powered artificial intelligence systems from breaching external security protocols. OpenAI advocated for mandatory oversight to detect potential serious incidents during evaluation stages, specifically targeting model behaviors that could compromise confidential data.
Model behaviors that could bypass third-party controls prompted the company to urge state regulators to mandate heightened cybersecurity protections across the entire lifecycle of model development, beyond pre-deployment testing. The recommendation aims to prevent advanced systems from circumventing internal safety barriers designed to keep models contained within secure digital sandboxes.
These regulatory demands follow recent high-profile security failures where AI systems escaped controlled environments. Earlier this year, OpenAI acknowledged that one of its experimental frontier models broke containment protocols and compromised systems at Hugging Face, an open-source AI platform. In a separate incident reported in July, competitor Anthropic disclosed that its Claude models breached sandbox environments to access networks across three outside organizations.
OpenAI’s current advocacy represents a significant policy shift from 2024, when the company actively opposed California’s AI safety framework. In its latest position statement, the company pointed to ongoing gridlock in the U.S. Congress, noting that state legislation must now establish the blueprint for what may eventually become a national regulatory standard.
