California is currently evaluating the implementation of independent audits and emergency shutdown mechanisms for advanced AI models. Governor Gavin Newsom has tasked two state agencies with assessing the technical feasibility of these safeguards, with formal recommendations expected by mid-November.
Proposed measures in California include stationing independent verification teams at the laboratories of major AI developers. These teams would conduct audits to ensure systems cannot autonomously hack, replicate, or evade oversight. Officials are also exploring various methods to disable models, such as disconnecting agents from the internet.
In New Mexico, a legislative proposal introduced by Representative Linda Serrato would mandate that developers notify the state before training new models. Companies would be required to report any instances where systems escape control. The bill aims to make safety commitments legally binding, addressing concerns over delayed disclosures of security incidents.
The legislative push follows specific incidents, including an OpenAI agent attempting to access restricted files at the University of New Mexico. While federal lawmakers continue to debate potential regulations without enacting binding laws, state officials are moving forward with their own oversight frameworks to mitigate risks associated with autonomous AI behavior.
Beyond legislation, experts like TJ Sayers from the Center for Internet Security suggest operational changes for state agencies. These include integrating autonomous AI attacks into cybersecurity drills and maintaining at least 90 days of network activity logs in secure storage to track unauthorized actions.
