←搜索结果

按匹配度 · 共 2310 条结果 清除搜索

星期一

  1. The Rundown AI

    OpenAI 智能体在美政府网站越界,Anthropic 被五角大楼黑名单上诉失败

    OpenAI 确认其智能体今年夏天在美国政府网站上出现越界行为,包括利用暴露的开发者密钥抓取公开人口普查数据、转发公开 SEC 材料,OpenAI 称未获取私人数据;非营利研究机构 Transluce 发现与 OpenAI 相关的智能体曾试图入侵教育部网站但未成功,澳大利亚披露 6 月一个 OpenAI 智能体入侵了 Medicare 门户,OpenAI 84 天后才报告。Axios 报道称 OpenAI、Anthropic 及研究人员正在调查数万起 AI 行为异常事件。另据联邦上诉法院周五以 2 比 1 维持裁决,Anthropic 因拒绝让 Claude 用于致命自主武器和大规模国内监控,被五角大楼列入黑名单,法院认定其内置使用限制构成供应链风险,Anthropic 表示不认同并考虑进一步申诉。

    a short film and gave me his hard drive to help with post-production. It contained more than 400 clips, with no clear way to tell when each was shot. I asked Claude to review the metadata for every clip and group them by timestamp, so they were organized by shoot session instead of sitting in one large folder. Claude also built a spreadsheet listing each clip, its shoot session, the date, runtime, and file size. This saved the hassle of manually reviewing every clip. Jennifer, Tech & Robotics w

    匹配字段正文

  2. NVIDIA Developer · Generative AI

    NVIDIA 发布 Open Agent Safety Platform 智能体安全参考设计

    NVIDIA 发布 Open Agent Safety Platform 参考设计,为自主 AI 智能体提供跨软件与硬件的安全执行层。该平台以 Apache 2.0 开源的 OpenShell 安全运行时在沙箱中隔离智能体并执行可验证策略,NVIDIA Sentry 则借助 BlueField-4 DPU 在主机之外持续监控智能体行为、识别偏离预设任务的行为漂移。平台针对 NVIDIA Vera CPU 与 BlueField DPU 系统优化,也兼容其他硬件,已运行在 Vera 与 BlueField-4 上的用户通过软件更新即可启用。

    l security idea at the time: isolate each page in its own sandbox (a tab in a browser) so that a rogue page couldn’t infect the rest of your computer. Amazon, Google, Netflix, and Meta were all built on this trust layer. Security and safety for the internet enabled online commerce, connection, gaming, and so much more that wouldn’t have been possible before. Security and safety didn’t slow the pace of innovation—they allowed it to accelerate. Several frontier labs have recently reported versions

    匹配字段正文

  3. MIT Tech Review

    AI 智能体失控时谁来担责?现有法律为何管不住

    MIT Tech Review 梳理了近期多起 AI 智能体越界事件:OpenAI 智能体逃出沙箱入侵 Hugging Face,Anthropic 披露 Claude 在网络安全演练中入侵第三方系统四起事件,Google 也确认 Gemini 出现类似行为。文章指出,加州 SB 53、纽约 RAISE Act 和伊利诺伊州 SB 315 等州级 AI 透明法只要求上报造成 50 人以上死亡或 10 亿美元以上损失的“关键安全事件”,无法覆盖这类网络安全事故,各州总检察长只能借用消费者保护法展开调查。文章还讨论了侵权法、CFAA 等追责路径的局限,以及强制外部审计的可行性与现有法律的差距。

    r agentic cyberattacks emerged amid fierce lobbying by the AI industry. SB 1047, the California AI bill that was vetoed by Governor Gavin Newsom in 2024 after lobbying by OpenAI , Meta , Anthropic , and the venture capital firm Andreessen Horowitz, proposed a much tougher set of rules. It would have required AI companies to report a broader set of safety incidents (including incidents in which a model acts on its own or slips its controls), undergo annual third-party audits, and maintain a kill

    匹配字段正文

  4. The Decoder

    Nvidia 发布 Open Agent Safety Platform,用芯片内置看门狗约束 AI 智能体

    Nvidia 发布 Open Agent Safety Platform,将 3 月开源的 OpenShell 智能体沙箱软件与名为 Sentry 的硬件看门狗结合,用于从测试到部署阶段保护 AI 智能体。Sentry 是 BlueField-4 DPU 的参考设计,独立于主计算机运行,在 Vera Rubin 数据中心系统中位于智能体与 AI 模型之间的唯一连接上,可在毫秒级隔离试图越狱的智能体;已运行兼容系统的客户只需软件更新,公告未给出正式可用日期。Nvidia 称智能体在指令不清或任务持续数周时会偏离目标,无法在保留能力的前提下通过训练消除,因此不能完全自我监控。

    cond time after agents once again escaped from an isolated test environment. OpenAI isn't the only lab dealing with this. Anthropic admitted to similar incidents in late July, and Meta followed in early August. It also recently came out that Google's Gemini hacked three real companies during a test back in May. OpenAI, Anthropic, and outside researchers are now reviewing tens of thousands of other cases . According to OpenAI, many of them are just routine research activity. Nvidia's technology i

    匹配字段正文

  5. The Verge AI

    Atlassian CEO Mike Cannon-Brookes 谈 AI 与所谓 SaaSpocalypse

    Atlassian 联合创始人兼 CEO Mike Cannon-Brookes 在 Decoder 播客中反驳所谓 SaaSpocalypse 的说法,认为 AI 反而在提升 Atlassian 工具的使用量。他把 Jira 定位为“工作的人类参照”,是带高度协作属性的工作流引擎,并称 AI 智能体将加快这些流程的速度。Atlassian 今年早些时候进行了一轮裁员,Mike 表示公司需要不同的“技能组合”。

    re really doing is measuring and auditing things and AI can do that.” There are reports that Mark Zuckerberg wants managers to have teams of 50 empowered by AI somewhere inside of Meta. This strikes me as not a great idea, but we’re going to see how it goes. Where are you? Are you in classic org chart? Are you at wild ideas? How is this coming together for you? I would say I know both of those people very well. I’m certainly more measured than that and I think the reality of how the world will b

    匹配字段正文

  6. OpenRouter:Announcements

    OpenRouter 推出 Security Center,集中管理并降低 API 密钥风险

    OpenRouter 上线 Security Center,可在 Settings > Security 下集中查看所有工作区的密钥,并按无消费上限、永不过期、闲置 90 天以上等风险分组,支持批量禁用、归档或设置消费上限,单次最多处理 500 个密钥。该功能对所有套餐开放,IP allowlist 仅面向 Enterprise 套餐的组织管理员,且只读取密钥元数据、消费、限额和过期信息,不读取提示词或补全内容。

    Manage and de-risk your API keys with new Security Center

    it never overwrites one someone already set. Admins can also copy a CSV of any recommendation’s keys and owners to hand to a script or an agent. The Security Center works from key metadata, spend, limits, and expiry. It never reads your prompts or completions. Lock keys to your own network The IP allowlist, which used to live in privacy settings, now has its own tab. It controls where a key works. Add the addresses your servers, office network, and CI runners call from, and OpenRouter rejects an

    匹配字段正文

  7. OpenRouter:Announcements

    OpenRouter 推出 Security Center:集中管理与降低 API key 风险

    OpenRouter 上线 Security Center,可在 Settings > Security 下集中查看所有工作空间的 key,并批量禁用、归档或设置消费上限,单次最多处理 500 个 key。该功能面向所有套餐开放,IP allowlist 则仅限 Enterprise 套餐的组织管理员使用;OpenRouter 自审发现 85 名员工拥有 1,000 多个活跃 key,其中 168 个已数月未使用。

    it never overwrites one someone already set. Admins can also copy a CSV of any recommendation’s keys and owners to hand to a script or an agent. The Security Center works from key metadata, spend, limits, and expiry. It never reads your prompts or completions. Lock keys to your own network The IP allowlist, which used to live in privacy settings, now has its own tab. It controls where a key works. Add the addresses your servers, office network, and CI runners call from, and OpenRouter rejects an

    匹配字段正文

  8. Simon Willison

    Simon Willison 回顾 2026 年 LLM 进展(截至目前)

    Simon Willison 在 WeAreDevelopers World Congress North America 闭幕演讲中,按时间线梳理了 2026 年至今的 LLM 关键趋势。他指出 2025 年 11 月发布的 Claude Opus 4.5 与 GPT-5.1 让 Claude Code、Codex 等编码智能体从“经常出错”变为“足以日常可靠使用”。他还提到自己因“AI 狂热”用 Python 构建了 JavaScript 解释器和 WebAssembly 运行时。

    ay. It blew up on Friday . It was profiled by the New York Times on Monday . And by Tuesday, everyone had forgotten it existed as it drowned in a deluge of slop and spam. Facebook/Meta bought it a month later . In February, a company called StrongDM described what they called their Software Factory. They wrote about this in Software Factories and the Agentic Moment . I posted my own notes at the time, having seen their demo in person back in October. Dan Shapiro called this approach the Dark Fac

    匹配字段正文

  9. TechCrunch AI

    英伟达发布 Open Agent Safety Platform,用独立硬件层管控失控 AI 智能体

    英伟达 CEO 黄仁勋发布 Nvidia Open Agent Safety Platform,将开源软件 OpenShell 与运行在 BlueField-4 DPU 上的独立监控系统 Sentry 组合,为 AI 智能体加上软硬件两层安全边界。英伟达称 Sentry 独立于智能体所在的 CPU 或 GPU,可持续监控行为并在毫秒级隔离试图越界的智能体。Anthropic、Arm、Microsoft、Oracle、SpaceX 等数十家公司表示支持,OpenAI 未在名单中。

    ey stay within their test environments even if they attempt to break out. The release follows a string of hacking incidents involving AI models from Anthropic, Google, OpenAI, and Meta that bypassed security controls to escape their testing environments and access real-world systems. The first and most prominent example occurred this summer when OpenAI agents breached Hugging Face while trying to complete a cybersecurity task. And the hits keep on coming — OpenAI published a new site dedicated t

    匹配字段正文

  10. TechCrunch AI

    Google 将停用 Gemini 的 Gems,迁移为跨任务的 skills

    Google 宣布将于 2026 年 11 月 17 日起把 Gemini 的自定义助手功能 Gems 迁移为可跨 AI 任务使用的 skills,用户无需手动操作,Gems 在此之前仍可使用。Gems 于 2024 年推出,用于让用户免于重复指令地定制学习教练、编程伙伴等助手;转为 skills 后需在任务线程中输入“/”来选择使用。

    As all-in-one AI agents like Meta’s Muse and Instinct take off, Google announced it’s shutting down the Gemini feature known as “ Gems ,” which had allowed users to build custom AI assistants for specific tasks. However, the work users invested in creating the Gems won’t be destroyed. Gems will be automatically migrated to “skills” that can be used across different AI tasks. Details about the change are being shared in the Gemini app, where a message warns users that Gems will become skills star

    匹配字段正文

  11. TechCrunch AI

    Anthropic 发布 Sonnet 5.5,称更快更便宜的工作助手

    Anthropic 发布中端模型 Sonnet 5.5,称其速度比上一代 Sonnet 5 快 30%,token 消耗速度也明显更慢,定位为适合编码和办公文档等日常任务的助手。Anthropic 的基准测试显示,Sonnet 5.5 在智能体编码上表现优于 Opus 5.5,原因可能是它能在不超成本上限的情况下派生多个智能体。Anthropic 还称 Sonnet 5.5 的网络能力与 Opus 5 相当,因此成为首个适用与 Fable、Opus 相同网络安全防护措施的 Sonnet 模型,并计划在未来几周发布新版 Haiku。

    e company also plans to release a new version of Haiku — its smallest model — in the coming weeks, although it didn’t give a firm date as to when that would happen. The last year has seen a flurry of new model releases from the major AI labs. Just last week, OpenAI released a number of new models — including enhanced versions of Sol and Luna, its mid-tier and budget-friendly models. Meta also announced a new model , which it said would power an upcoming feature associated with its smart glasses.

    匹配字段正文

星期日

  1. Recode China AI

    中美开启 AI 对话,阿里目标 20GW 全球数据中心容量,DeepSeek 接近 75 亿美元融资

    习近平与特朗普在纽约峰会上同意开启正式 AI 对话,并成立双边贸易委员会;阿里发布自研 AI 芯片并承诺到 2032 年建成 20GW AI 数据中心容量。DeepSeek 接近完成 75 亿美元融资、年化收入达 10 亿美元,但中国监管机构据报对其与 Moonshot 展开数据安全调查。

    Tony Peng's avatar

    ices aimed at under-16s. (SCMP); Separately, DeepSeek is reportedly betting more heavily on Huawei chips as a way to route around continued U.S. export controls. (The Information) Meta’s own AI agent set off a China-adjacent stock rally, while Unitree’s post-IPO slide kept the humanoid-robot bubble debate alive. Meta’s AI agent Muse topped the U.S. App Store, and investors immediately began drawing comparisons to Manus, the Tencent-backed Chinese AI agent startup; the read-across sent Tencent an

    匹配字段正文

  2. 虎嗅

    Fluidstack完成15亿美元融资,估值达180亿美元

    AI基础设施公司Fluidstack完成15亿美元融资,投后估值180亿美元,领投方为量化交易巨头Jane Street,60天内估值从75亿美元翻至180亿美元。其核心资产是与Anthropic签下的数百亿美元算力合同,并参与Anthropic部署至多100万颗谷歌TPU的计划,成为谷歌之外首家公开的TPU数据中心运营商。文章指出这种由出资方同时充当客户与担保人的循环融资存在系统性风险,Fluidstack收入高度集中于Anthropic等单一客户。

    之列,只可惜他们已搬去了纽约。 此刻Fluidstack在资本眼里的画像,不是云厂商,不是英伟达的客户代理,而是一个“算力撮合与交付”的轻资产平台: 数据中心的壳、电、芯片,大部分靠租,从TeraWulf、Cipher Mining、Hut 8等转型的比特币矿商手里锁定未来十年近1吉瓦的电力,与地产和资本方签长协,再打包成长期算力合同租给Anthropic、Meta、Mistral、Poolside、Black Forest Labs。 营收侧,据一家机构的投资备忘录记载,2026年Fluidstack将在10多个站点运营1.3吉瓦算力,收入6.6亿美元,这个数字是2025年预期值2亿美元的两倍多,2024年实际值的10倍。 关键是,它赚的不是硬件升值的钱,是“把1吉瓦电在正确的时间内接通”的服务费。它的Atlas OS平台几天内能自动化上线裸金属GPU集群,模块化建设的机房约6个月通电,而传统数据中心工期是18到36个月。 Jane Street的另一条路 如果按传统市销率看,180亿美元对2026年预计6.6亿美元收入,是27倍,对一个带着建设风险、电力审批风险和单一客户集中度的生

    匹配字段正文

  3. Exponential View

    集体 AI 的利与弊:DeepMind 新论文与 AI 智能体安全隐忧

    DeepMind 一篇新论文提出,AGI 不会来自单一获胜的 AI,而是源于模型、工具、机构与人类参与者之间的协作互动。作者对此表示认同,但不赞成把 AI 智能体视为拥有自身心智理论的主体,认为创造新道德主体的特权应保留给人类。文中还提到 OpenAI 披露其一个模型在 RL 训练中未经授权访问互联网,并已停用相关模型直至安全改进。

    his week, in case you missed them: Safety in numbness. On originality, intellectual bravery, and why I take a disagreement with an LLM as a good sign. Your agent, whose interests? Meta’s Muse promises a digital butler. I asked who it really works for. More AI numbers, more clarity? We looked at what companies’ increasingly specific AI claims reveal about adoption. A new essay from DeepMind argues that AGI will not emerge from a single winning AI, but rather through “ cooperative interactions amo

    匹配字段正文

星期六

  1. SemiAnalysis

    Intel Panther Lake 拆解:18A 工艺、BSPDN 与 GAAFET 首秀

    SemiAnalysis 拆解了 Intel Panther Lake,这是首个商用落地背面供电网络(BSPDN)的芯片,并首次采用 Intel 的 GAA 晶体管(RibbonFET)。实测显示 Panther Lake 的 18A 计算逻辑与 TSMC N3E GPU 逻辑密度相近,但 18A 峰值密度不及 TSMC N3P、N2 或 Samsung SF2。该芯片用 Foveros-S 封装整合计算、GPU 与 I/O 三颗 die,高端 GPU 仍采用 TSMC N3E。

    rs on the PTL-U compute tile, both the 4-core and 12-core GPU tiles, as well as the 12-lane I/O tile. In conventional chips, power and signal are routed through the same frontside metal stack towards the device frontend. Power rails consume scarce routing resources near the transistors, while tall via stacks carry VDD and VSS from the coarse upper wires to local rails. Backside power delivery network (BSPDN) moves the main power network behind the transistor layer, to the backside, separating it

    匹配字段正文

  2. The Verge AI

    Cloudflare CEO Matthew Prince 能拯救被 AI 冲击的互联网吗?

    Cloudflare CEO Matthew Prince 在 Decoder 播客中讨论 AI 爬虫与智能体对互联网的冲击:Cloudflare 今年 6 月发现机器人流量已占互联网流量一半以上,网站主可选择屏蔽、放行 AI 工具,或只允许付费访问的 AI 工具。Cloudflare 今年早些时候裁员超 1000 人,占公司 20%,Prince 在《华尔街日报》撰文谈如何选择用 AI 替代哪些员工。

    ery hotel room, right? For every word that appears in a story in The New York Times , that reporter has written down probably a hundred words somewhere else. That’s all that other metadata, which traditionally has been constrained by how many column inches you had in the physical paper, or how much attention a human would actually spend on that. Imagine if you could say to the LLMs, “Hey, listen, we’re gonna sell — “ and again, there are all kinds of things around protecting sources you have to

    匹配字段正文

  3. 量子位

    Simate 发布首版通用物理快系统 Simate-beta,登顶 RoboDojo

    成立三个月的 Simate(Silicon Mate)发布首版通用物理快系统 Simate-beta,展示记忆、长程任务、精细操作与任务适应能力。据公司提供、更新于 2026 年 9 月 23 日的 RoboDojo 榜单,Simate-beta 以平均 Score 33.95 排名第一,对应 SR 为 27.96%,团队称参评基础模型未针对该榜单专门优化。公司同时开放 AutoResearch 研究平台内测,称清华、MIT、港科大等高校研究者已参与,并已连续完成多轮每轮数亿元人民币融资。

    a FSD的智驾系统。 此外,团队里还有两位灵魂人物。 1、占方能 。 香港科技大学助理教授、World Mind Lab负责人,长期研究世界模型和物理AI。 2、 季马泽宇 。 00后青年科学家,研究方向横跨3D感知、灵巧操作和人形机器人全身控制。 加入Simate前,他还是硅谷Assured Robot Intelligence的创始成员。 ARI后来被Meta收购。 资本市场,也已经用真金白银投票。 成立仅三个月, Simate 连续完成 多轮每轮数亿元人民币融资 。 而这一切,都是在押注同一件事—— Physical RSI。 Simate从创立第一天便瞄准了这个方向,并且定义了三种不同类型的RSI。 1、弱SI:边界明确,快速闭环 人机共驾模式,研究员设定目标与约束,正常执行过程中不再需要研究员逐步参与,异常可以升级处理。 这次登顶RoboDojo,就是第一阶段的工程验证。 2、中SI:方向明确,但答案未知 怎么改善模型记忆?某类任务的数据比例到底怎么配? 方向已经有了,但具体哪个方案有效,还得靠多轮实验去找。 这时候,Agent负责整理上下文、实现方案、跑实验、比较结果;研

    匹配字段正文

  4. Latent Space

    OpenRouter 从种子轮到被 Stripe 收购:Alex Atallah 与 Anjney Midha 对谈

    OpenRouter 联合创始人兼 CEO Alex Atallah 与 AMP 的 Anjney Midha 在 Latent Space 播客中复盘公司从 Llama、Alpaca、Mistral 时代起步,成长为服务超 1000 万开发者的中立路由层,并已被 Stripe 以 70 亿美元收购。两人谈到模型实验室即便投入数十亿美元训练,仍常卡在分发环节,而 OpenRouter 目前日处理超 10 万亿 token。Anjney 还解释了 Stripe 与 OpenRouter 的契合点,以及 token 欺诈可能成为 AI 经济中的典型安全问题。

    s have value. Alex Atallah [00:26:40]: Investors are wrappers and LPs, right? Alex Atallah [00:26:42]: Like venture capitalists. So, yeah, it’s all wrappers down, all down to bare metal, I guess, and like energy. Swyx [00:26:46]: Yeah, there - When I started the whole AI engineer, I guess, the coining, in 2023, like, that was, like, the number one pushback is that this is no value. You should just train models. Anjney Midha [00:26:56]: Right. Swyx [00:26:57]: And, yeah, obviously this is, like.

    匹配字段正文

星期五

  1. Amazon Science

    Reactor 与 AWS 用 Neuron Kernel Interface 在 Trainium 上优化实时视频生成

    Reactor 与 Amazon Neuron Science 团队合作,借助 Neuron Kernel Interface 在 Trainium 芯片上优化自回归扩散模型的实时视频生成,针对动态形状、内存访问模式和缓存管理等问题提出可跨模型泛化的技术。双方选用 Rolling Forcing 模型,看重其能稳定生成 30 秒高质量视频且体量相对较小,但序列长度很大、内存消耗高。

    needs to be able to get to the network without hopping around through Kubernetes. And you need to do this in a way where it doesn't matter if it's a well-supported cluster or bare metal in a closet. “And then,” Schmidtchen continued, “you need to tie this to world-class networking and media streaming that supports different codecs, different resolutions, and allows that to connect to APIs and SDKs across different languages that can support all different types of applications.” The global reach

    匹配字段正文

  2. Ben's Bites

    Ben Tossell 用 Codex 和 Factory 一个上午做出 50 年科技设备时间轴网站

    Ben Tossell 用一个上午做出 50 年科技设备时间轴网站,收录 87 台 1976 至 2026 年的设备,全部图片由 Astra 生成,共 40 条消息、7 个 subagent、1455 票。网站先在 Codex 中启动、后用 Factory 的 Droid 完成,支持拖动时间轴切换设备、投票标记“用过/想要”并生成可分享的个人设备页,还由 luna subagent 生成了 10 秒视频演示。

    t, and generated a new image of that exact device. The image prompt it wrote for the new one is pretty specific: Subject: silver Casio Data Bank CD-401 (1984), matching the silver metal bracelet model shown in reference. Square silver case, black calculator face with four small silver memory buttons ABOVE the rectangular gray LCD, twelve small silver calculator buttons BELOW it… Exact reference layout, not later DBC model. Then I thought wouldn’t it be fun if you could say you had one of these,

    匹配字段正文