California's SB 813 voluntary AI safety framework

California's legislature approved SB 813 on September 4, establishing a voluntary framework of Independent Verification Organizations to conduct third-party safety assessments of frontier AI models before release.

Critics argue the framework is already outdated: frontier models have grown so large that other AI models are required to analyze them, and a METR investigation into the OpenAI Hugging Face hack consumed $400,000 in AI compute paid for by OpenAI. One analyst wrote that such laws 'become outdated almost immediately'.

Averi announced an IP-safe eval methodology in which the eval company never accesses model weights and the model creator receives no information about the evaluation

Dean Ball argued for legislated AI audit mechanisms, drawing analogy to financial auditing, following a false viral claim about OpenAI using 'neuralese'