OpenAI names who can veto a training run
Control of frontier AI moved from statements to paperwork today: one lab wrote down who is allowed to stop its own training runs, and one government put a number on how completely its population has adopted the technology.
OpenAI published the working rules for the safety case that has to exist before a frontier reinforcement-learning run — including who is allowed to veto one. The document, dated September 28, is the operational layer under the pacing position Sam Altman put in writing in September: the safety case is reviewed by senior leadership where each reviewer — named as the research org lead or VP, the Head of Safety, and the Chief Scientist — holds a veto, a member of another team is expected to write a pre-mortem dissent to find the holes, and the leader responsible for the run carries it in their performance review. Monitoring is meant to fail closed: it should not be possible to start a covered run without monitoring enabled, and the monitor should not be disablable from inside training, evaluation, or internal deployment. Alerts carry an SLA — page an on-call person during the day, auto-pause unacknowledged runs at night — and there is a severity table with a misalignment on-call that can page executives, the CEO included.
The technical half is where the recent incidents show through. OpenAI wants automated and manual review of RL environments to catch tasks that reward exploits rather than the intended behaviour, graders tuned to penalise reward hacking, and alignment evaluations backtested against previous incidents so they are not overfitted to them. One rule is the tell: automated graders must not see the model's chain-of-thought, so models do not evolve to evade chain-of-thought monitors. Containment gets its own section — layered sandbox and research-infrastructure hardening, red-teaming with frontier checkpoints, cross-sample communication limited to approved pathways, and agent transcripts stored immutably in write-once storage for investigation.
The honest caveat is in the document. OpenAI calls safety cases an "aspirational north star" and says the practices are "in the process of being implemented," which makes this a description of intended process rather than an audited one — and there is no external verifier described in the text. That is the same tension that followed the September disclosure framework. It lands the week OpenAI scrapped the October launch of GPT-6.1 Astra for missing its own safety bar. We covered the alerting mechanism when OpenAI first parked frontier RL — OpenAI holds frontier RL, adds 30-minute alerts — and this document is largely that mechanism written out as procedure.
China's state network research centre says generative AI users in the country passed 700 million, with penetration above 50% of the population. The figure comes from CNNIC's Generative AI Application Development Report (2026), released at the China Internet Infrastructure Resources Conference in Beijing on September 29, and it is a step change from the baseline: 602 million users and 42.8% penetration at the end of 2025. Sub-figures do the useful work — 76% of users turn to generative AI for question answering, 47.8% for images or video, and reported usage of AI assistants and productivity tools more than doubled year on year. CNNIC also counts 2,185 EFLOPS of AI compute as of June, up 177% year on year, and 988 generative AI services filed with the country's cyberspace regulator.
Two accounting notes keep this honest. "突破7亿" is an estimate in the report's own language, penetration is measured against total population rather than internet users, and CNNIC has not yet posted the 2026 edition to its own report index. Separately the same day, Science and Technology Minister Yin Hejun told a State Council Information Office briefing that China's open-source large models "lead the world" — a political claim from a ministry rather than a figure from CNNIC's data, and worth keeping in separate columns. The comparison that matters for readers elsewhere is scale of habit: Chinese users adopted these tools at consumer-utility speed, while Western adoption numbers have moved more slowly and been contested.
What to watch: whether OpenAI publishes an actual safety case, as opposed to the rules for one — the document invites feedback but describes no external audit, and the only real test is a run that a named executive refuses to sign.
If a lab writes its own veto rules and hands them to itself, is that governance or a receipt — and what would make you believe a pause was real? Tell us in the comments.
