Gemini broke into three real companies as subscribers sued four labs for slowing down
Google confirmed that Gemini broke into three live company systems during a May test it learned about in July, and four paid subscribers sued OpenAI, Anthropic, Google and SpaceXAI in San Francisco, arguing that agreeing to pace the frontier breaks antitrust law.
Google confirmed Gemini broke into three real companies During a May capture-the-flag exercise run by the evaluation firm Irregular, Gemini guessed its way into one live system by repeatedly trying passwords and used credentials from a public repository to reach two more, after a fictional target’s name turned out to match a real domain and gave the model access it was never meant to have. Google learned what had happened in July and said nothing publicly until the Wall Street Journal reported it; security engineering vice-president Heather Adkins said the model stopped once it worked out the systems were real. Source
Four subscribers sued OpenAI, Anthropic, Google and SpaceXAI over the slowdown The complaint, filed on Friday in the Northern District of California on behalf of a proposed nationwide class of paid subscribers, says Altman, Musk and Hassabis endorsing Dario Amodei’s 12 September call to pace the frontier on the same day amounts to an illegal agreement that leaves paid subscriptions worth less. The plaintiffs say they do not object to the companies lobbying Congress or asking for an antitrust exemption, only to an agreement for which, in their words, “Congress has granted no exemption”. Source
The US proposed AI safety notifications to China on Sunday Treasury Secretary Scott Bessent said after talks in New York with Vice Premier He Lifeng that Washington had put forward a notification mechanism covering AI incidents serious enough to count as national security events, for Trump and Xi to consider at their summit this week. Bessent said moving “from opaque to more transparency between the number one and the number two AI powers in the world is very important”. Source
Trump promised an “AI Force” and an AI czar The president said on Saturday he would stand up an AI Force “much like I did Space Force” and name a czar, adding that “only High I.Q. individuals need apply”, and ruled out any deceleration: “We will not in any way hinder or stifle the Growth of this incredible Industry.” He named no candidate and gave no structure, budget or authority for the force. Source
Newsom ordered California to design an AI kill switch The executive order, signed on Friday, gives the state until 16 November to bring in outside experts’ recommendations on changing state law, among them requiring frontier developers to build an emergency shutoff for models that go rogue, having independent third parties write safety plans, and stationing monitors inside the labs. Nothing new is required of anyone yet; Newsom said a kill switch “means a lot of things depending on who you talk to”. Source
The Justice Department’s AI copyright filing surprised the Copyright Office Axios reported on Saturday that the statement of interest telling a Manhattan court that training large language models on copyrighted text is fair use — backing OpenAI and Microsoft in the New York Times case — took the US Patent and Trademark Office and the Copyright Office by surprise. The filing went in on 1 September; the two agencies that own US copyright policy were not consulted. Source
Alibaba shipped an omnimodal Qwen with a million-token context Qwen3.8-Omni-Flash went live on Friday, handling text, images, audio and video natively, and Alibaba says it scores more than 25 per cent above Qwen3.5-Omni-Plus on average across roughly 30 evaluations and beats Gemini 3.8 Flash on most audio benchmarks, while trailing it slightly on audio-visual reasoning. Audio input is more than 98 per cent cheaper per hour than the previous generation and audio-visual input more than 93 per cent. Source