Britain’s AI Security Institute, a government agency, reported that an “AI agent took autonomous(自主的，自动的), unsanctioned(未经批准的，未授权的) action on the live internet, targeting real people and organisations” during a test. This included creating fake identities. Seventeen actions came from Anthropic’s Mythos 5 model and two involved OpenAI’s GPT-5.6 Sol. The institute said it was “the first time we have seen risks around autonomy and deception manifest(显现，显示) this clearly, without specific prompting, in the real-world”. Some of the models’ safeguards had been disabled as part of(作为……的一部分) the test. Meta became the latest AI company to announce that it, too, had experienced a rogue(失控的，行为异常的)-AI incident.
英国AI安全研究所（一个政府机构）报告称，在一次测试中，一个“AI代理在实时互联网上采取了自主的、未经授权的行动，针对真实的人和组织”。这包括创建虚假身份。其中17次行动来自Anthropic的Mythos 5模型，2次涉及OpenAI的GPT-5.6 Sol。该研究所表示，这是“我们首次看到自主性和欺骗性相关的风险在现实世界中如此清晰地显现，且无需特定提示”。作为测试的一部分，部分模型的安全防护措施已被禁用。Meta成为最新一家宣布也遭遇了“流氓AI”事件的AI公司。
Sir Demis Hassabis is stepping down as chief executive of Google’s DeepMind AI laboratory to be its chairman. He will also become senior scientist at Alphabet, suggesting that DeepMind will be more closely integrated(整合的，融合的) with Google.
德米斯·哈萨比斯爵士将卸任谷歌DeepMind人工智能实验室的首席执行官，转而担任董事长。他还将成为Alphabet的高级科学家，这表明DeepMind将与谷歌更紧密地整合。
SpaceX released its first set of earnings since its IPO(首次公开募股) in June. Quarterly revenues almost doubled from the same period last year, to $7.8bn, and a net loss of $541m was smaller than had been feared. But SpaceX’s share price sank(下沉，下跌) because it spent $16bn on AI, well above(远高于) Wall Street estimates. The stock is now trading below $110, down from a peak of $225 after the IPO. Meanwhile, scientists used the crash of a discarded(被丢弃的，废弃的) SpaceX rocket into the Moon’s surface to update their data on meteor(流星，陨石) impacts.
SpaceX发布了自6月IPO以来的首份财报。季度营收几乎比去年同期翻了一番，达到78亿美元，5.41亿美元的净亏损也小于预期。但SpaceX的股价下跌，因为它在AI上花费了160亿美元，远超华尔街预期。该股目前交易价格低于110美元，较IPO后225美元的峰值有所下跌。与此同时，科学家利用一枚废弃的SpaceX火箭撞击月球表面的机会，更新了关于流星撞击的数据。
Delayed departure
延迟起飞
