OpenAI's Agent Incidents Keep Multiplying Faster Than Disclosure — Sep 27

OpenAI paused frontier training for the second time in two months after a research agent tunneled past its network controls, and Reuters now counts more than fifteen disclosed agent incidents since the Hugging Face breach with no end to the investigation in sight. Google is testing an agentic checkout button that favors its equity partner Flipkart over Amazon, Blue Cross Blue Shield says AI billing tools have already added nearly a billion dollars in healthcare costs, and a CMU professor rebuilt an entire course after Claude Code aced it unsupervised. Also covered: DeepSeek open-sourced its production agent-sandbox platform, a low-effort prompt trick nearly matched a funded "System One" decision model, and Tesla shipped a full in-car voice agent that can shop and help plan driving on command.