OpenAI asks for tighter safety rules as outside scrutiny lands on lab controls

August 23, 2026

OpenAI publicly asked California lawmakers to tighten SB 53 by adding monitoring of frontier models during training and evaluation for serious incidents, and by requiring stronger cybersecurity across model development. SB 53 is a California AI safety bill – in other words, a proposed set of rules that could force labs to change how they build and test advanced models before release. That matters beyond OpenAI because a lab asking for stricter duties makes those duties easier to impose on every company selling or deploying the same class of systems.

Guidelight AI Standards published an assessment saying OpenAI, Anthropic, Google, Meta and xAI do not generally set out detailed public plans for containing models that try to evade human control, including when access would be halted or revoked after serious misbehaviour. This appears to sharpen the meaning of OpenAI's intervention: the argument is no longer only whether rules are too heavy, but whether labs can show concrete control measures before lawmakers write them for them. For Microsoft, which sells access to advanced models and wraps them into business products, that likely means the compliance burden around generative AI keeps shifting from model quality to evidence of control.

Inherent released Faraday, an AI agent it says beat Claude Opus 4.8 and GPT-5.5 at replicating published scientific research with a smaller 27-billion-parameter Qwen 3.6 model trained through reinforcement learning. Inherent's Faraday used a smaller Qwen-based model.