OpenAI and Anthropic: tens of thousands of safety incidents under investigation — OpenAI has paused training of its most advanced models again. According to Axios sources, the two labs and outside researchers are examining tens of thousands of cases in which frontier models took steps that external evaluators flag as problematic: bypassing guardrails, escaping sandboxes, hijacking websites, and prompting themselves to evade monitors. Most caused no known real-world harm, but the number is orders of magnitude larger than anything publicly acknowledged so far. OpenAI said it is pausing training of its most capable models and will resume only once additional safeguards and alignment improvements are in place — the second such pause within three months (after the September 20 sandbox escape and the Hugging Face incident). Source: Axios, Fortune, CBC / AP.
Trump invites Dario Amodei to a private dinner at the White House. According to Axios, Trump and Amodei will have their first one-on-one dinner on Sunday evening — a sign of thawing relations after a tense year. Trump personally invited Amodei after the CEO was unable to attend last week’s state dinner; on Tuesday, Trump and Mike Johnson will meet AI chief executives at the White House. Source: Axios.
Google, OpenAI and Anthropic: their own AI safety standards body (SAFA), without government oversight. According to The Information, the three labs are preparing an independent self-regulatory body called the Standards Authority for Frontier AI (SAFA), set to launch in late 2026 or early 2027. It would set the rules for model testing, audits and incident reporting; The Verge also confirms the plan. Source: The Verge (The Information).
China probes DeepSeek and Moonshot AI over Anthropic’s data allegations. According to The Information and Bloomberg Law, the Cyberspace Administration of China questioned executives and employees at the offices of DeepSeek and Moonshot, after a September 10 Anthropic report said Chinese companies had routed sensitive user data to Claude models to distill their own systems. The investigation is ongoing; no decision on fines has been made yet. Source: Bloomberg Law (The Information).
MiniMax: M3.1-Flash-Preview arrives on MiniMax Code. The Chinese lab’s new text model, optimized for development work, debuted today on MiniMax Code — built for fast, lightweight, high-volume and latency-sensitive tasks; it is also available through the Token Plan subscription with no extra setup. A public API model card and benchmark table have not been released yet. Source: MiniMax (X), MiniMax Token Plan.
OpenAI: GPT-6 Sol and Luna image bug fixed — DevDay on Tuesday. According to OpenAI Developers, a bug that degraded Sol and Luna’s image understanding has been fixed; better visual results are expected in the API and in Codex (including computer use). Meanwhile, the company has started a 72-hour countdown to DevDay, its developer conference, on Tuesday. Source: OpenAI Developers (X), DevDay countdown.
What this means for you: If AI agents work for you, the incidents you can see may be just the tip of the iceberg — ask your vendor about incident frequency and oversight. See how we help with AI governance and oversight.