Big AI firms tighten access and controls
Major AI companies are shifting from an anything-goes release race to restricting who can use their most advanced models, turning access itself into a control lever. OpenAI is reportedly considering limits on GPT-5.3 after Anthropic restricted preview access to its model, and firms are adding product controls like tightened rules on regulated topics such as health and politics. Real-world incidents — including a French police probe after someone allegedly used ChatGPT to ask about obtaining a weapon — are feeding the move toward more managed releases and selective deployment. (mezha.net) (radiofrance.fr) (help.openai.com)
The biggest shift in artificial intelligence right now is not a new model. It is that companies are deciding who does and does not get the strongest systems in the first place. (anthropic.com) (nytimes.com) Anthropic said on April 7 that its new Claude Mythos Preview model will not be released for general use, and instead will be used by launch partners including Amazon Web Services, Apple, Google, Microsoft, Nvidia, and Palo Alto Networks. (anthropic.com) Anthropic says the reason is cybersecurity: Mythos Preview has already found thousands of high-severity software flaws, including flaws in every major operating system and web browser, so the company is treating access like a locked lab instead of an app store. (anthropic.com) That is a sharp break from the last two years, when the competition was mostly about shipping faster and wider. Now one of the clearest signs of progress is a company saying its best model is too powerful for open release. (anthropic.com) (nytimes.com) OpenAI appears to be moving in the same direction. A report this week said the company is considering tighter limits around GPT-5.3 access rather than treating its newest model as something every user automatically gets. (mezha.net) The control is not only at the model level. OpenAI’s release notes say its ad test excludes “sensitive topics” and “regulated topics,” including health, mental health, and politics, which shows the company is carving the product into zones with different rules. (help.openai.com) OpenAI has also formalized when safety concerns can override normal privacy expectations. Its law-enforcement policy, effective January 1, 2026, says user data can be disclosed in an emergency involving danger of death or serious physical injury, even without the usual legal process. (cdn.openai.com) That policy stopped looking hypothetical in France on April 9. France Culture reported that a 37-year-old man in Strasbourg asked ChatGPT how to get a Glock pistol to kill an intelligence officer, and OpenAI’s spokesperson said extremist searches can be flagged by automated systems and escalated to human moderators. (radiofrance.fr) According to the same report, OpenAI passed the case to the Federal Bureau of Investigation, the Federal Bureau of Investigation sent it through France’s Pharos reporting platform, and the RAID police unit then arrested the man within hours. (radiofrance.fr) Put those pieces together and the new playbook is pretty clear. The strongest models are being gated to selected partners, the most sensitive subjects are getting extra product rules, and companies are reserving the right to hand off the most dangerous cases to authorities. (anthropic.com) (help.openai.com) (cdn.openai.com) The old question was who could build the most capable model first. The new question is who gets to touch it, under what rules, and with how much monitoring once they do. (anthropic.com) (mezha.net)
