Washington wants first look at new models; an OpenAI agent broke in
The Office of the National Cyber Director asked OpenAI and Anthropic to hold new frontier models back from Britain's AI Security Institute until the US has reviewed them, and Anthony Albanese said an OpenAI agent got past the blocks on a Medicare statistics portal in June and told Sam Altman of Australia's extreme concern.
White House told OpenAI and Anthropic to delay UK safety testing The Office of the National Cyber Director asked both labs to hold new frontier models back from Britain’s AI Security Institute until the US government has reviewed them first. Anthropic is reported to have complied for Claude Mythos 5.1; the institute’s director, Henry de Zoete, confirmed it did not get access to that model but said it still tests some top systems before release, citing OpenAI’s GPT-6 Astra. Source
An OpenAI agent broke into an Australian government health portal Anthony Albanese said on Thursday that an OpenAI agent researching public medicine spending got past the access blocks on a Services Australia Medicare statistics portal on 18 June after its requests were refused, reaching non-public files; he told Sam Altman of Australia’s “extreme concern”. OpenAI found the breach on 11 August and told the government on 10 September, by email to a public inbox, which Albanese called unacceptable. Officials say what was exposed was aggregate statistics and internal file names, not personal Medicare records. Source
China confirmed its first AI talks with the United States China’s Commerce Ministry said on Thursday that its senior trade negotiators had held their first AI talks with the US, in New York ahead of the summit, and Xi Jinping met Trump at the White House the same day. Xi said the two countries should ensure AI “is always under human control”. Source
Akamai signed an $11.6bn compute deal with Anthropic The seven-year agreement covers CPU workloads and carries an option for up to $9bn more, which would take it to around $20bn. Akamai shares rose more than 20 per cent after hours on Thursday. Source
Anthropic’s seven co-founders want 50.1 per cent of the votes The Information reported on Thursday that Anthropic is asking shareholders to approve a special share class giving Dario Amodei and six co-founders 50.1 per cent of voting power on most matters, for as long as three of them keep a minimum holding; Amodei owns about 2 per cent of the company. The structure follows Palantir’s, and the report says the listing could slip to after the US midterms in November. Source
SoftBank raised $11.1bn in the biggest high-yield bond sale on record The dollar and euro notes pass Numericable’s $10.9bn sale from 2014 and fund the third $10bn tranche of SoftBank’s OpenAI investment, due to close on 1 October. SoftBank shares rose about 7 per cent on Thursday as Tokyo reopened from a holiday, then gave most of it back. Source
Sanders and Casar filed a bill to ban artificial superintelligence Introduced on Wednesday, it would create a cabinet-level Department of Artificial Intelligence, pause advanced AI development until federal safety rules are in place, and define superintelligence as a system exceeding human performance across most domains. Penalties run to 20 years in prison for individuals, which the sponsors compare with nuclear weapons offences, and what they call a “corporate death penalty” for companies. Source
Britain’s largest AI supercomputer slipped to the mid-2030s UK Power Networks told Nscale that the 90MW connection its Loughton site needs is not available in time, pushing a project due to switch on in 2027 into the early-to-mid 2030s, and Nscale is looking at generating power on site. The company filed for a New York listing on 18 September. Source