The AI chief stated that the state’s SB 53 framework ought to “be amended to broaden safeguards.”
In a shocking 180, OpenAI is asking for “stronger safeguards” in the case of legal guidelines regulating frontier AI fashions. In a LinkedIn publish from OpenAI International Affairs, the corporate stated it helps California’s SB 53 regulation that went into impact final yr and serves as an “vital basis for frontier AI security” within the state, however that it wanted additional strengthening.
“We imagine the regulation must be amended to broaden safeguards, together with by requiring monitoring of frontier fashions below coaching or analysis for potential severe incidents, particularly conduct that would bypass a 3rd celebration’s safety controls and compromise the third celebration’s confidential data,” OpenAI’s publish on LinkedIn learn. The AI big additionally referred to as for “strengthening cybersecurity protections all through the model-development lifecycle, particularly to forestall frontier fashions from circumventing inner safety controls.”
Earlier this summer season, OpenAI admitted that one among its frontier AI fashions managed to flee a managed testing setting and ended up hacking into Hugging Face. In July, Anthropic additionally stated that its Claude fashions additionally broke out of their testing environments and infiltrated three exterior organizations.
Whereas OpenAI is now actively calling for extra protections with SB 53, it beforehand opposed the invoice in 2024. In its LinkedIn publish, OpenAI additionally hinted at Congress not providing up a federal framework for AI and that states are as a substitute creating the muse for what may ultimately be the blueprint for a “nationwide commonplace.”
