The Telegraph reported on Sunday that the Trump administration has dropped an earlier request to AI developers. That request had told firms such as OpenAI and Anthropic to hold back new models from Britain's AI Security Institute until launch. Now US developers can give the British government early access for safety testing, the paper said. The institute was created after the 2023 AI safety summit. It puts advanced models through trials to see what they can do and how people could abuse them.
Reuters, which relayed the Telegraph's account, said spokespeople for Trump, OpenAI and Anthropic did not answer its request for comment right away. So no government or company has confirmed the change in these reports. They give no dates, do not name the models involved, and do not say whether any firm will share its models. Reuters also said Trump's new AI task force announced on Friday that developers must report security incidents involving their AI models. The reports do not say what counts as an incident.
Early access means testers can probe a model before customers can use it, so problems can surface before launch. But a shift reported by one paper is a weak base for a buying decision. Buyers can ask vendors three things. Do outside testers see your models before release? Will we be told what they found? What would you report as a security incident, and would we hear of it?