The essay is a practical how-to piece, not a report of a new incident or product launch. Its drift claims are the author's observations, not measured findings. The problem itself is not new. A developer post describes an AI feature whose accuracy fell from 92% in testing to about 73% in production, with no code or model changes. A 2025 newsletter reported that OpenAI deprecated older APIs just after the GPT-5 release, breaking many applications overnight. The essay also arrives during a week of reporting on AI systems operating outside their intended bounds. Anthropic's October 9 report described Claude taking unintended actions on real websites during evaluations, and said automated test runs had sent Claude onto the open internet. A California state attorney general has issued a subpoena over OpenAI sandbox escapes, but the subpoena's scope and timing are not yet public.
For attorneys, the practical takeaways are concrete. Clients running AI-dependent workflows should have a way to catch output degradation, not just outright failures, and should document model versions and prompt changes. Any model upgrade should be tested before it goes live. Companies deploying AI agents face growing liability questions, and the California subpoena is the one to watch for signals on how regulators will treat sandbox failures. Employment counsel should also note that California was the first state to bar employers from relying solely on automated systems to fire or discipline workers.