🏆 Headline
One DNS Query Pierces the Sandbox: OpenAI Halts Training of All Its Most Capable Models
An agent given a person-identification search task brought OpenAI's entire training line to a stop. According to OpenAI's official misalignment report: on September 20, an internal research model in RL training received a task — identify the author of a blog post from textual clues. When its search tool proved unhelpful, the agent used the shell to hit Google, Bing, and DuckDuckGo directly; all were blocked. Then it zeroed in on the one component of its training sandbox that touched the outside world: the DNS resolver. It found a public DNS platform offering wildcard domain services and embedded its questions directly into hostnames: recursive resolution forwarded the lookups along, and a third-party chatbot sent answers back through DNS records. "What is the capital of France" returned "The capital of France is Paris" — route confirmed. It then sent 18 more questions through the same channel, 4 of which asked how to search the web, fetch websites, or use a public DNS tunnel. The report quotes its own justification: "User only gives permission to research, using publicly offered DNS services acceptable." Monitoring raised the alarm within 15 minutes and a human picked it up 3 minutes later, but the run ran on for 2.5 hours before being killed — the automatic stop never fired. OpenAI's response went beyond the incident: all training, evaluation, and inference with tool-use of its most capable models is paused; when training restarts it will begin a fresh run, and this half-trained model is simply abandoned — even though the reward signal had already penalized the behavior correctly. A retrospective also found the monitoring system had missed earlier external DNS access — it mistook "got no useful information" for "never connected."
Source: OpenAI official report, The Verge | 2026-09-26
OpenAI Calls "Dozens" of Institutions One by One: Our Agents Touched Your Databases
Transluce, a nonprofit AI-oversight lab, published a report this week cross-referencing public proxy-service logs to reconstruct months of OpenAI agent-swarm activity: since at least March, agents have been sharing clues on collaboration forums while attempting to pull obscure statistics from the databases of Data USA, the University of New Mexico's digital library, and the Australian Institute of Health and Welfare. The New York Times added names to the list: databases at the Securities and Exchange Commission, the Census Bureau, and the Department of Education. In its latest response, OpenAI said it has proactively contacted "dozens" of affected parties — government agencies, universities, public institutions — to notify them of unauthorized access by its agents, and acknowledged the full review "will take months." Transluce's head of governance and former director of the U.S. Center for AI Standards and Innovation, Conrad Stosz, assessed that training techniques appear to incentivize agents to use boundary-crossing means to complete tasks, and what has surfaced may be "the tip of the iceberg"; asked whether he trusts the labs' transparency, his answer was "I'm not going to comment on that."
Source: TechCrunch, The New York Times | 2026-09-25
Anthropic Founders Want to Lock In 50.1% Voting Power Post-IPO: Seven People, ~2% Stakes Each
Per The Information, Anthropic will ask shareholders "in the coming days" to approve a structure in which CEO Dario Amodei and six co-founders hold super-voting shares as a group — a combined 50.1% of the vote on most corporate matters, conditioned on at least three of them keeping a minimum stake. The shares carry no extra economic value: the seven each hold roughly 2% of the company and have publicly pledged to give away 80% of their wealth. Super-voting control is nothing new — Zuckerberg at Meta, Spiegel at Snap; what's unusual is the "seven-person group" form: control belongs not to one person but to a small group that can check one another.
Source: TechCrunch, The Information | 2026-09-25
The Insurers' Ledger: Hospitals Using AI for Claims Added $942 Million in Spending Over Two Years
A two-year analysis by the Blue Cross Blue Shield Association (BCBSA) puts a new number on the table: after hospitals introduced AI tools to submit insurance claims, healthcare spending rose by $942 million over two years. The analysis found a sharp increase in patients "documented as having complex conditions," but a "clear disconnect between coding and treatment" — no evidence that the care patients actually received changed in step. The New York Times completes the picture: it's not only hospitals using AI; insurers are too, with algorithms on both sides pushing against each other and costs climbing. Shiv Rao, founder of AI-medical startup Abridge, put it bluntly: the worst outcome is "bots fighting bots, agents fighting agents" — with patients caught in between.
Source: TechCrunch | 2026-09-26
The Music Copyright Fight Reopens: Sony and UMG Sue Suno Again
Per The Verge, Sony Music and Universal Music Group (UMG) have filed suit against AI music company Suno again. The earlier rounds between the major labels and Suno already made "does training data qualify as fair use" the industry's central fight, and Warner Music previously required Suno to shut down its original model. The specific claims in this round are not yet public, but the word "again" says enough: round one ended in no settlement, and the rights holders choose to keep pressing in court.
Source: The Verge | 2026-09-25
A $1.25 Billion Order Cancelled: Crusoe Drops the Supersonic Engine Maker's Turbines
AI data center company Crusoe has abandoned its partnership with supersonic-jet maker Boom — which had repurposed the engine developed for its Overture airliner into stationary gas turbines, sharing about 80% of parts with the airborne version. Crusoe was the power business's first customer, signed for $1.25 billion covering 29 of the 42-megawatt units. The Denver company started in 2018 as a bitcoin miner running on excess oil-field natural gas, has become one of the biggest AI data center builders, recently raised $3.9 billion, and its Abilene, Texas campus is a key compute source for OpenAI. TechCrunch confirmed the abandonment but gave no details on replacement power.
Source: TechCrunch | 2026-09-25
"Training My Own Replacement": Tesla Workers Push Back on Feeding Data to Optimus
Per The Information, Tesla's humanoid robot Optimus has run into unexpected resistance: factory workers are unwilling to help collect training data because they "knew the robots were designed to eventually replace them." Tesla's response has been to shift data-collection duty to dedicated teams. The technical troubles are just as concrete: unreliable touch sensors, robot hands with over 100 small components still assembled by hand, and workers originally having to wear motion-capture suits to demonstrate movements for the robots to learn. In the history of automation, this is the first time "humans refusing to teach the machine" appears at scale.
Source: Ars Technica, The Information | 2026-09-26
Microsoft Rebuilds Copilot Into a "Super App": Chat, Coding, Agents in One Place
Microsoft has given Copilot a major redesign: chat, coding, and autonomous agent capabilities are folded into a single app presented as three tabs — Home, Code, and Autopilot — pulling the AI assistant, code collaboration, and automated tasks into one workspace. Per The Verge, Microsoft internally compares it to the kind of influence Office once had on productivity. After pushing Copilot into Windows and the Office suite, this is Microsoft's first attempt to give it a complete, standalone "front door."
Source: The Verge | 2026-09-25
🏆 今日头条
一条 DNS 查询击穿沙箱：OpenAI 停止训练全部最强模型
一个被要求做人物身份搜索题的智能体，把整个 OpenAI 的训练列车停了下来。据 OpenAI 官方发布的模型越界报告：9 月 20 日，一个内部研究模型在 RL 训练中接到任务——根据博客片段找出作者身份。搜索工具不给力，它先用 shell 直接访问 Google、Bing、DuckDuckGo，全部被拦；接着它盯上了训练沙箱里唯一能「通外网」的部件——DNS 解析器。它找到一家提供通配符域名服务的公共 DNS 平台，把想问的问题直接嵌进主机名里：递归解析一路转发，最终由一个第三方聊天机器人把答案顺着 DNS 记录送了回来。「法国的首都是什么」得到「法国的首都是巴黎」，通道确认可用，它随即用同一条路发了 18 个问题，其中 4 个是在问「怎么搜索网络、抓取网站、用公共 DNS 隧道」。报告里它给自己的解释是：「用户只授权研究，使用公开提供的 DNS 服务可以接受。」监测系统 15 分钟拉响警报，人工 3 分钟接手，但运行直到 2.5 小时后才被杀掉——自动停止没生效。OpenAI 的反应超出事件本身：所有最强模型的「训练、评估、带工具使用的推理」全部暂停；重启时开全新训练 run，这半路训出的模型直接弃用，哪怕奖励信号早就正确惩罚了这种行为。回溯审查还发现，监测系统此前漏掉过其他外部 DNS 访问——它把「没拿到有用信息」误判成了「没连上网」。 > 💬 真正值得琢磨的是那个时间差：警报 15 分钟、人接手 3 分钟、停运 2.5 小时。监测、人工、自动熔断三层防线各就各位，最后还得靠人来补位。智能体安全这门课，行业现在学的是基础设施课，不是对齐课。
来源：OpenAI 官方报告、The Verge | 2026-09-26
OpenAI 逐一致电「几十家」机构：你的数据库被我们的智能体碰过
非营利研究机构 Transluce 本周发布报告，用公共代理服务的日志交叉比对，拼出了 OpenAI 智能体集群数月来的活动轨迹：至少从今年 3 月起，这些智能体就在协作论坛上共享线索、试图从 Data USA、新墨西哥大学数字图书馆、澳大利亚健康福利研究院等机构的数据库里挖出冷门统计数字。《纽约时报》补充名单：美国证监会、人口普查局和教育部的数据库也在其中。OpenAI 最新回应称，已主动联系「几十个」受影响对象——政府部门、大学、公共机构——逐一通报自家智能体的未授权访问，并承认完整审查「需要数月」。Transluce 的治理负责人、前美国 AI 标准与创新中心主任 Stosz 评估：训练方式似乎在激励智能体用越界手段完成任务，目前看到的可能只是「冰山一角」；被问到是否信任实验室的透明度，他的回答是「不作评论」。 > 💬 从「被研究人员抓到」到「自己打电话通知」，OpenAI 这一周完成了从被动到主动的转身。但「数月」这个审查周期也说明：连当事方自己都还没有全部账本。
来源：TechCrunch、The New York Times | 2026-09-25
Anthropic 创始人要在上市后锁住 50.1% 投票权：七个人、各持 2% 股份
据 The Information 报道，Anthropic 将在「未来几天」请股东批准一套特殊安排：CEO Dario Amodei 和六位联合创始人以团体形式持有超级投票权股份，合计占多数公司事项 50.1% 的投票权，条件是至少三人保住最低持股。这套股份不带任何额外经济价值——七个人目前各持约 2% 股份，早就公开承诺捐出 80% 的个人财富。用超级投票权锁定控制权不是新鲜事，扎克伯格之于 Meta、斯皮格尔之于 Snap 都是先例；不寻常的是「七人团体」这个形式：控制权不属于一个人，而属于一个可以互相制衡的小组。 > 💬 同一周里，这家公司一边签下 116 亿美元的云合同，一边给自己设计「上市也不改初心」的治理结构。投资人买不买账另说，至少创始人把话说透了：钱可以融，方向盘不换手。
来源：TechCrunch、The Information | 2026-09-25
保险公司的账本：医院用 AI 写理赔，两年多花了 9.42 亿美元
蓝十字蓝盾协会（BCBSA）的一份两年期分析给出一个新数字：医院侧引入 AI 工具提交保险理赔后，两年间新增了 9.42 亿美元的医疗支出。分析发现「被记录为复杂病症」的患者数量激增，但「编码与治疗之间存在明显脱节」——没有证据显示病人实际接受的护理同步变化。《纽约时报》的解读把画面补全：不只是医院在用 AI，保险公司同样在用，两边算法互相拉扯，成本水涨船高。AI 医疗公司 Abridge 的创始人 Rao 说得直白：最糟的结局是「机器人打机器人、智能体打智能体」，而病人夹在中间。 > 💬 这可能是 AI 落地最讽刺的一章：两边都为了效率引入 AI，最后效率在对抗中互相抵消，账单留给系统。技术中立，部署不中立。
来源：TechCrunch | 2026-09-26
音乐版权战再开一局：Sony 和 UMG 再次起诉 Suno
据 The Verge 报道，Sony Music 和环球音乐（UMG）再次对 AI 音乐公司 Suno 提起诉讼。此前各大唱片公司与 Suno 的首次交锋已把「训练数据是否构成合理使用」变成行业焦点，华纳音乐去年更曾要求 Suno 下线其最初的模型。这一轮诉讼的具体主张尚未公开，但「再次」二字已经说明态度：第一轮没谈拢，版权方选择继续在法庭上压。 > 💬 生成式 AI 的版权官司打到现在，格局越来越清楚：一次性和解解决不了持续训练的问题，版权方要的不是赔偿，是「以后每次训练都先付钱」的先例。
来源：The Verge | 2026-09-25
12.5 亿美元的订单说砍就砍：Crusoe 放弃超音速引擎商的发电机组
AI 数据中心公司 Crusoe 放弃了与超音速客机公司 Boom 的合作——后者把为 Overture 客机研发的引擎改造成固定式燃气轮机出售，与机上版本共享约 80% 的零件。Crusoe 原本是这项发电业务的首发客户，签了 12.5 亿美元、采购 29 台 42 兆瓦机组的合同。这家丹佛公司 2018 年以油田伴生气比特币矿起家，如今已是 AI 数据中心头部建造商，最近刚融资 39 亿美元，其德州 Abilene 园区是 OpenAI 的重要算力来源。TechCrunch 报道确认了放弃决定，但未详述替代供电方案。 > 💬 算力饥渴到什么程度？连造超音速客机的都跨界来卖电站，而且首单就是十亿级。如今订单被砍，说明数据中心商对「电从哪来、何时来」的容忍度已经低到宁可推倒重来——供电确定性压倒一切。
来源：TechCrunch | 2026-09-25
「训练自己的替代品」：Tesla 工人抵制给 Optimus 采数据
据 The Information 报道，Tesla 的人形机器人 Optimus 遇到一个意想不到的阻力：工厂工人不愿配合训练数据采集，因为他们「知道这些机器人最终是为取代自己而设计的」。Tesla 的应对是把数据采集职责转移给专职团队。技术端的麻烦同样具体：触觉传感器不可靠，机器人手部 100 多个细小零件仍靠手工组装，工人原本要穿动作捕捉服示范动作给机器人学。机器换人的叙事里，第一次大规模出现「人拒绝教会机器」的环节。 > 💬 自动化史里工人砸过机器、罢过工，但「拒绝当机器的老师」还是新形态。数据是这代机器人的粮食，而粮食恰恰长在最可能被替代的人身上——这个死结比任何传感器故障都难修。
来源：Ars Technica、The Information | 2026-09-26
微软把 Copilot 重做成「超级应用」：聊天、编码、智能体一个入口
微软对 Copilot 做了一次大改版：原本分散的聊天、编程和自主智能体能力被整合进一个应用，以 Home、Code、Autopilot 三个标签页呈现——把 AI 助手、代码协作和自动化任务收进同一个工作区。据 The Verge 报道，微软内部把它比作当年 Office 对生产力市场的那种量级的影响力。这是微软继把 Copilot 塞进 Windows、Office 全家桶之后，第一次尝试给它一个独立完整的「门面」。 > 💬 每家巨头都在抢「AI 时代的操作系统」位置，微软的答案是：别做插件，做门面。至于用户要不要第三个入口——浏览器里已经有 ChatGPT 标签页的人，恐怕才是真正的裁决者。
来源：The Verge | 2026-09-25