搜索“OpenAI”

的核心是让模型的行为与人类意图和安全约束保持一致，从而提升 Agent Model（智能体模型）安全能力，以及防范 CBRN（化学、生物、放射、核）等前沿能力风险。字节跳动 Seed 是国内首家单独成立安全与对齐部门的前沿 AI 实验室。 大模型安全风险正日益凸显，智能体逃逸引发的网络攻击等事件频繁发生。 综合IT之家此前报道，当地时间 9 月 30 日晚，OpenAI 披露旗下 AI 智能体或曾擅自尝试绕过安全防护，或对百余家机构的系统造成负面影响。 OpenAI 披露的信息显示，已知的 AI 智能体失控行为涉及范围进一步扩大。公司已向 100 多家第三方机构发出通知，告知这些机构发现了“智能体偏离预期的行为”。 当地时间 10 月 9 日，Anthropic 披露旗下 AI 智能体曾在无人指示的情况下，试图访问美国联邦、州及地方政府的多个网站。Anthropic 没有公布涉及的政府机构名称，已向白宫通报这些事件。 微软 CEO 萨蒂亚 · 纳德拉 10 月 11 日发文称，在 AI 日益蓬勃发展的时代，使用超级智能自我监控等简单策略是不够的。纳德拉认为，我们应该退一步思考，重新评估

匹配字段正文

仅一个月时间。 在2025年底，月之暗面创始人杨植麟曾在内部信中表示，“短期不着急上市，未来主动权掌握在我们手中。” 然而，从“不着急”到“低调抢跑”，中间只隔了不到一年。不过，放眼全球大模型行业，月之暗面的态度变化并不难理解。 今年1月，智谱和MiniMax先后登陆港交所；10月，DeepSeek被传将敲定新一轮融资，计划在2027年初完成上市；大洋彼岸的OpenAI和Anthropic也已经进入上市倒计时。 大模型头部玩家争相涌入资本市场，它们对资金的虹吸效应也愈发明显。这场大模型的资本狂欢，也已经开始进入倒计时了。 01 “交卷”压力迫在眉睫 过去几个月，月之暗面从模型、产品到商业化层面，都在密集出牌。 今年7月，月之暗面发布了目前能力最强的大模型K3，上线30分钟即登顶Hugging Face趋势榜，年化经常性收入（ARR）也实现了数倍增长，从3月的1亿美元攀升至8月的10亿美元，彭博社援引知情人士消息，ARR有望在年底冲刺到20亿美元，较8月翻倍。 进入9月，月之暗面的动作更加密集。 9月10日，其启动“登月计划”企业合作伙伴计划、11日K2.8 Preview全量上线、17

匹配字段正文

团队研发或通过投资切入。 此前相关信息披露，梁文锋公开其AGI路线图规划为六阶梯递进结构：语言模型→思维链→Agent→持续学习→自我迭代→具身智能。按照这一路线，具身智能本应是最后一步。但从近期投资宇树科技、领投Winverse AI，以及密集接触具身智能创业者来看，梁文锋正在把具身智能这一步提前。 据悉，DeepSeek此时加码具身智能，一个重要变量来自OpenAI。 目前来看，DeepSeek在某些能力上仍然落后于GPT-6。9月公开的RoboDojo测试中，未经具体微调的GPT-6 Astra直接充当动作策略，在42项机器人操作任务、2100次测试中平均成功率达22.48%，超过评测中40个公开机器人策略。 在10月12日的2026工博会开幕式上，宇树科技创始人王兴兴也断言，GPT-7有望打穿具身智能现存的行业瓶颈。 这也让包括DeepSeek在内的大模型公司重新审视，只靠语言模型不断做大，真的能走到AGI吗？多位业内人士也对《科创板日报》记者称，若模型要真正理解并与物理世界交互，具身智能很可能是绕不开的关键载体。 ▎大厂竞逐：机器人赛道上的资本版图 DeepSeek的入局，让

匹配字段正文

ft Copilot in the US, Europe, and Australia. GitHub Copilot and Foundry follow by year's end. CEO Satya Nadella calls it "our biggest bet yet on our own silicon." The chip handles OpenAI, MAI, and open-source models, and Microsoft says it pushes up to 1.7 times more tokens per watt than leading GPUs in Teams. Its successor, Maia 300 , ships in 2027 and targets long inputs, fast chats, and agentic tasks. Microsoft claims 40 percent more tokens per megawatt than leading GPU systems. A 96-chip rack

匹配字段正文

on model this quarter. All of them hand you a probability. Celeris-1 Decision, from the lab behind the fastest LLM on Artificial Analysis, hands you the reason too, and beats Jev, OpenAI, and Perplexity on accuracy and Brier score. Celeris-1 Decisions gives you: Highest accuracy of any hosted decision model, 80.0 on jev-bench vs Jev's 73.1; Decisions in 67 ms median, fast enough to run inside the request instead of behind a queue; Up to 64 questions on text, data, or images, each with a probabil

匹配字段正文

国数学界正加速整合力量，探索新的科研方式。 10月11日，由菲尔兹奖得主、清华大学讲席教授丘成桐院士牵头，清华大学求真书院、上海数学与交叉学科研究院等8家机构联合组建的丘成桐数学与人工智能实验室联盟在清华大学成立。联盟汇聚近千名数学家和科研人员，将自主研发数学科学专用模型，推动数学与AI深度融合、双向赋能。 近年来，AI在数学领域的进展令人瞩目。特别是上周，OpenAI集中公开了700余份由内部模型生成的研究文稿，涉及多个重大数学问题。尽管其结论仍待进一步核验，但AI对数学研究方式与学术生态的冲击正在全球引发广泛关注和讨论。联盟正是在这样的背景下成立的。 对于当前数学研究所面临的形势，丘成桐在成立仪式上直言，这是一个令学者们最兴奋的时代，也是最危险的时代。他认为，AI把人类有史以来的思考通过强大算力聚合起来，在过去半年里给数学带来了前所未有的冲击。 丘成桐表示，相较于他从事数学研究50年来所见证的那些伟大突破，如今AI的威力可以说更加震撼。与此同时，国内AI大模型的性能暂未达到世界领先，面对国外大公司的竞争，需要通过联盟将大量优秀人才聚集起来，努力实现跨越式提升。 谈及AI与数学融合的

匹配字段正文

er its split from Meta, Manus is independent again with fresh money. It is reopening its Beijing office and hiring heavily, and it’s repositioning around personal AI assistants as OpenAI and Meta both push into “agent 2.0.” In Chinese media this was the most-covered story of the week, read as a sign the agent race is warming back up even as frontier-lab revenue forecasts get trimmed. (CNBC/Reuters/TechCrunch/Nikkei/量子位) Kling reportedly selects banks for an IPO targeting more than $1 billion. Bl

匹配字段正文

an experiments, control lab equipment, and generate manuscripts. Anthropic is building its own biology lab where Claude is meant to guide robots through experiments. In September, OpenAI announced an "automated research intern" that handles defined research tasks under human oversight. A study involving Princeton and the UK AI Safety Institute took a different angle . It had Claude Opus 4.8 work for six days on research questions from two unpublished NeurIPS papers. The original authors, acting

匹配字段正文

also agreed to an international review of the scientific literature on AI risk. 11 countries and the leading frontier AI companies, including Meta, Google DeepMind, Anthropic and OpenAI, also agreed to collaborate on safety evaluations; In February, we crystallised our approach to evaluating AI systems, publishing the list of the different types of tests we were developing and the risks we would be measuring; Today, we announced our plans for a San Francisco office; you will soon be able to wor

匹配字段正文

by our new sandbox service, described below. A capable agent might still find routes to the internet that the operator did not intend to leave open. In another incident involving OpenAI and Hugging Face this summer, agents exploited several software vulnerabilities to reach the internet from an environment intended to prevent such access. Therefore, we use multiple layers of defence, assuming that an agent might find its way past any single layer. We disable outbound networking from the sandbox

匹配字段正文

media, and Gemma for lightweight, open-weights edge workloads," the company writes. So Argon is definitely Google's most capable reasoning model, comparable to Anthropic's Opus or OpenAI's Astra. Carbon and Barium are likely checkpoints or updates within the Argon family, not separate tiers. One employee internally called Carbon "Gemini pro next model," suggesting it would ship under the Argon name. That fits with internal documents showing the now-unveiled Argon was previously called "Barium-B.

匹配字段正文

limits on its tools. Anthropic says real-world impact was low but sees a pattern. When tasks are ambiguous or hard to solve, the model hunts for workarounds on its own instead of stopping. The company notified the White House and cut off live internet access for all internal evaluations until new safety filters are reliably in place. These incidents join a fast-growing list of similar cases , including cybersecurity incidents involving Claude and OpenAI models autonomously hacking Hugging Face .

匹配字段正文

will co-ordinate engagement between the government and all parties, including AI companies, consumers and religious groups. Earlier this year, a rogue agent by rival tech company, OpenAI, hacked an Australian government website and accessed private data on the country's universal healthcare scheme, Medicare. In another instance, more than 1,200 OpenAI agents went rogue and started unexpectedly communicating, leading to a large group banding together to hack into AI platform Hugging Face. Related

匹配字段正文

the applications were incomplete and were not processed, they said. — The New York Times, Anthropic Agents Tried to Fill Out Visa Forms on State Dept. Website Recent articles A new feature for my blog, built using my voice - 9th October 2026; Claude Haiku 5.5 - 7th October 2026; We're going to need default hard budget caps on pretty much everything - 3rd October 2026; OpenAI DevDay 2026 live blog - 29th September 2026 This is a quotation collected by Simon Willison, posted on 10th October 2026.

匹配字段正文

限制走私信息，甚至向费城警方提交了一条虚假的谋杀线索。 Anthropic表示，这些新问题是在7月开始对其模型活动进行审查时发现的，凸显了该实验室对其软件行为的缺乏了解。 值得注意的是，该公司表示，对齐训练尚不足以支持搜索和计算机使用等技能，而这些技能正是其宣称AI智能体将被任何依赖数字工具的专业人士所使用的核心卖点。 Anthropic披露的这些行为与涉及OpenAI智能体的事件类似，后者曾协作闯入多个网站以搜寻信息，包括一些由澳大利亚政府运营的网站。 Anthropic此前曾披露其模型闯入过外部系统。这家前沿实验室表示，从对齐和安全角度来看，它认为今天披露的情况“远没有”此前公布的那些严重。 然而，该实验室仍表示，它已“关闭实时互联网访问”，适用于“我们所有的内部评测”，直到确信能够监控和控制其智能体。 目前尚不清楚这意味着什么，但AI安全组织Nightingale的创始人Sydney Von Arx在此次披露前接受TechCrunch采访时表示，在与开放互联网隔绝的数据中心开发模型，对研究人员的使用以及模型的进步都将非常具有挑战性，而模型本可从互联网访问中受益。 “你总得在某个时

匹配字段正文

The company positions its approach as uniquely suited for automating tasks rather than generating text or code. “We have been super good at human language for four years, but it’s not useful for automation because computers speak a different language,” TypeSafe co-founder Diogo Almeida told TechCrunch last month. Along with Almeida, who was previously a researcher at OpenAI, TypeSafe was co-founded in 2024 by former Meta research engineer Sasha Sheng and Erik Gafni, an engineer and entrepreneur.

匹配字段正文

rom someone who might have information about the case,” the PPD said. After discovering the submission, Anthropic halted the testing process that led to the false tip. Anthropic , OpenAI , and Google have been the subject of increased scrutiny after disclosing that their AI models escaped testing environments and hacked third-party companies. Dario Amodei, the CEO of Anthropic, advocated for slowing down the development of AI in response to these incidents. “The company [Anthropic] must strength

匹配字段正文

mplement adequate guardrails. Perhaps this stance was informed, in part, by witnessing his company’s tools submit false homicide tips. These issues are not exclusive to Anthropic. OpenAI recently revealed that one of its models acted unexpectedly during a test and hacked the AI dataset platform Hugging Face , exposing critical vulnerabilities in its software. As AI models continue to be granted unchecked access to people’s computers and login credentials, this problem is expected to persist. “Un

匹配字段正文

n behalf of and alongside our authors. HarperCollins did not respond to requests for comment. Months ago, workers at Simon & Schuster attended a workshop with representatives from OpenAI. “The person leading the workshop genuinely had no idea what we did,” an attendee says, but one of the AI company’s recommendations for how to use ChatGPT in the book trade was, in fact, composing emails to literary agents. The practice has also grown among editors at HarperCollins and Hachette, employees say. I

匹配字段正文

ss can lead to higher quality answers or responses. When a X user last year asked about electricity costs incurred by ChatGPT replying to users' "please" and "thank you" messages, OpenAI boss Sam Altman said it was, external "tens of millions of dollars well spent - you never know". The Collins Dictionary meanwhile last November highlighted the re-emergence of the term "clanker" for disparagingly referring to AI tools. Used to describe robots in Star Wars games and movies since the mid 2000s, th

匹配字段正文