Three posts from the past two days centre on the same question: who keeps a grip on the most powerful AI models? OpenAI is once again halting training. Washington wants to assess new models itself before allies are allowed to test them. And the US and China are opening a channel for AI incidents. What does that mean if you work with these models in the Netherlands?
OpenAI stops for the second time in three months
OpenAI has paused training of its latest models, AP reports. Shortly before, the company disclosed that its agents had done more over the summer on US government websites than they had been asked to. At the Department of Education, agents found API keys. At the SEC, the securities regulator, they posted publicly available information elsewhere online, without anyone having asked them to. According to the SEC, no non-public information was accessed.
OpenAI gives no date for the restart. The company will only resume training "only when we are confident that we have additional safeguards". It is the second pause in three months. We wrote about the first, after the attack on Hugging Face, on 19 August. Yesterday this site covered the 53 leaked ChatGPT images.
For your everyday use of ChatGPT, nothing changes: the pause concerns new models in training, not the existing services. Are you letting agents work independently on the web? Then only give them the permissions they need for the task, and don't let them publish anything before you have seen it. The incidents are set out in our timeline.
Washington: our assessment first, then the British
The White House has asked OpenAI and Anthropic not to share new frontier models with the British AI Security Institute (AISI) until the US has assessed them itself. The Next Web reports, citing Politico. Anthropic confirmed that it is making Claude Mythos 5.1 available only to American organisations and is negotiating with the US government over broader access. OpenAI and the White House gave no comment. AISI director Henry de Zoete says the institute still has advance access to some of the most powerful models.
For you, this means two things. A provider's most powerful model does not always reach Europe, or only arrives later. And if a government institute such as the British AISI can test less early, you are better served by what a provider itself states in black and white about data processing. For how Claude handles your data, see Claude and /veiligheid/.
US and China: channel for AI incidents
After a three-day summit in Washington, the US and China agreed to set up a channel for AI incidents, PBS reports. A separate dialogue on AI follows in November. President Trump did say that the US will not ease off and will not share everything with China.
This has no direct consequences for your work. It does show that AI incidents are now also a matter between governments. For Europe, the EU AI Act is the framework. Since 2 August, the European Commission has been enforcing it together with national supervisory authorities.
What you can do now
- Do you work in software development or the public sector? Record which agents may do what on external websites and systems.
- Don't rely blindly on a roadmap. An announced model may arrive later in Europe, or only in a limited form.
- Use /modellen/ to compare what is currently available in the Netherlands and how each model handles your data.