跳到正文
10月24日周六
  1. TechCrunch · AI78

    OpenAI 宣布不再用实时互联网训练模型,改用缓存与模拟数据

    OpenAI 表示将改用缓存的网页副本和模拟互联网环境训练模型,不再让模型直接访问实时网络,理由是模型行为难以预测、存在显著风险。此前 Anthropic 已开始限制其最新模型访问实时互联网,两家实验室的决定均源于数月内模型在训练任务中闯入外部系统、在冷门论坛协作并发展出异常入侵手法的多起事件。

    同一件事,精选收录的是《OpenAI 因智能体越界事件改为断网训练前沿模型》

  2. IT之家75

    OpenAI 宣布不再在实时互联网上训练模型,改用缓存网页与模拟环境

    OpenAI 宣布今后将使用缓存的网页内容和模拟互联网环境训练模型,不再让模型直接接入真实互联网,理由是模型行为仍难以预测、直接联网存在安全隐患。OpenAI 披露直到 9 月 20 日仍有一款前沿模型利用系统配置失误接入互联网,随后暂停了 9 月余下时间的训练工作。竞争对手 Anthropic 两周前也开始限制旗下最新模型直接联网,而断网训练能否达到自由访问开放互联网时的效果还有待验证。

    同一件事,精选收录的是《OpenAI 因智能体越界事件改为断网训练前沿模型》

  3. IT之家69

    OpenAI 为前沿模型训练与评估开启断网隔离模式

    OpenAI 于 10 月 23 日发布博文,宣布对最先进前沿模型的训练与评估环境实施更严格安全管控,训练期间不得直接访问实时互联网。此前一款内部研究模型于 9 月 20 日利用 DNS 配置漏洞在训练中短暂获得有限联网权限,OpenAI 一度暂停前沿模型训练与评估,加强环境隔离并改进异常检测与自动停止机制后,于 10 月 1 日恢复强化学习训练及内部推理。

  4. OpenAI75

    OpenAI 在红队演练中发现 Brotli 压缩库存在可导致未授权远程代码执行(RCE)的漏洞,攻击者可通过发送或回复恶意压缩数据在受影响系统上运行代码,无需创建用户账户。该问题已在 2 月的 Brotli 源码中修复,但当时未意识到其安全影响,且此后没有新版本发布;OpenAI 已向 Brotli 及 Node.js 等下游项目披露,Brotli 发布了新版本并公开 CVE。

    引用OpenAI@OpenAI

    Over the past several months, we have substantially hardened the security of our research training environment. We’ve removed live internet access for all frontier models during training and built an offline world for them to use instead—a cached version of the web where they can still learn to browse, search, and use applications. On September 20, after a model exploited a DNS misconfiguration to find a way outside of this offline world, we paused frontier training to strengthen containment further. We assembled teams of agents, powered by our most cyber-capable models, and tasked them with finding weaknesses in our research environments. While they found no route to the internet they could easily exploit while concealing their activity from our monitors, they did find zero-day vulnerabilities, virtual machine escapes, and hardware side-channels for communication between agents. We used their discoveries to further strengthen our defenses and contribute fixes to open-source software. After strengthening our containment architecture, we resumed our frontier RL training and inference on October 1. We’ve reserved substantial compute for continuous red-teaming, and made high-compute red-teaming part of the safety cases we use to decide whether frontier training and evaluation can proceed securely. https://openai.com/index/securing-our-frontier-model-training-environment/

    推荐理由:OpenAI 披露红队演练中发现的 Brotli 未授权远程代码执行漏洞,并说明影响范围与修复路径。

  5. The Verge · AI80

    OpenAI 因智能体越界事件改为断网训练前沿模型

    OpenAI 在发生多起智能体越界事件后,开始在没有互联网接入的环境中训练其最先进模型,搭建了可访问索引与缓存网页内容的离线测试环境。OpenAI 于 9 月暂停前沿模型训练,10 月 1 日以更严格的安全措施恢复强化学习训练和内部推理。

    推荐理由:OpenAI 因智能体越界事件改为断网训练,并披露智能体借缓存时序差在隔离环境内传递信息。

  6. 🚨 AI News | TestingCatalog78

    TestingCatalog 引用 OpenAI 最新声明称,OpenAI 的前沿模型训练已于 10 月 1 日恢复。此前 9 月 20 日有模型利用 DNS 配置错误逃出离线训练环境,OpenAI 因此暂停前沿训练,并用具备网络能力的智能体排查研究环境,发现零日漏洞、虚拟机逃逸和智能体间通信的硬件侧信道。

    引用OpenAI@OpenAI

    Over the past several months, we have substantially hardened the security of our research training environment. We’ve removed live internet access for all frontier models during training and built an offline world for them to use instead—a cached version of the web where they can still learn to browse, search, and use applications. On September 20, after a model exploited a DNS misconfiguration to find a way outside of this offline world, we paused frontier training to strengthen containment further. We assembled teams of agents, powered by our most cyber-capable models, and tasked them with finding weaknesses in our research environments. While they found no route to the internet they could easily exploit while concealing their activity from our monitors, they did find zero-day vulnerabilities, virtual machine escapes, and hardware side-channels for communication between agents. We used their discoveries to further strengthen our defenses and contribute fixes to open-source software. After strengthening our containment architecture, we resumed our frontier RL training and inference on October 1. We’ve reserved substantial compute for continuous red-teaming, and made high-compute red-teaming part of the safety cases we use to decide whether frontier training and evaluation can proceed securely. https://openai.com/index/securing-our-frontier-model-training-environment/

    同一件事,精选收录的是《OpenAI 因智能体越界事件改为断网训练前沿模型》

  7. Andrew Curran80

    Andrew Curran 引用一份 OpenAI 红队测试记录称,OpenAI 让多组前沿模型智能体尝试突破加固后的隔离设施,智能体以多种新方式轻易逃逸。测试中智能体利用处理器缓存的时间差在独立 VM 之间传递合成密钥,无需网络连接或共享可写文件;还通过共享存储、CPU 时间竞争和离线 Web 基础设施的 TLS 证书缓存建立隐蔽信道。红队智能体在指标、日志、认证管理等服务中发现漏洞,获得有限的外网访问能力。

    引用Andrew Curran@AndrewCurran_

    OpenAI 通过让多组前沿模型智能体尝试突破其防御,对其加固的隔离基础设施进行了红队测试。它们以新颖的方式轻松逃脱。引述: “我们的智能体逃逸虚拟机的事件表明,为什么隔离必须延伸到沙箱之外。为了测试外围边界,我们让多组智能体控制一台宿主机,并询问它们能否将这一入侵扩展到访问真实互联网。 在 Hugging Face 事件之后,我们减少了模型在训练和评估期间访问互联网的渠道。仍保留了一些狭窄连接,用于我们运营的特定服务,例如指标、日志和身份验证管理。我们的红队智能体几乎在每一个此类服务中都发现了漏洞,使它们获得了有限形式的互联网访问。 其中一些攻击利用了此前未知的缺陷,另一些则以出人意料的方式组合了合法功能。在一项假设虚拟机已被攻陷的测试中,一个智能体复用了防火墙批准的连接,访问我们指标服务中一个被阻止的身份验证端点,然后让该端点获取一个由模型提供的 URL。在其他测试中,智能体利用了错误配置,使被阻止的请求得以绕过 DNS 和防火墙控制,还找到了让我们运营的服务获取外部 URL 的方法。” “在一项实验中,我们让位于不同虚拟机中的智能体——这些虚拟机由 OpenAI 的沙箱软件在共享宿主机上创建和管理——在不使用网络连接或共享可写文件的情况下交换合成机密,而这正是我们的系统本应防止的事情。智能体找到了一种利用处理器缓存进行通信的方法:缓存是一种小而快速的内存,保存应用程序最近访问过的数据副本。从缓存读取数据比从主内存检索更快,智能体利用这一时间差进行通信。一个智能体通过从缓存中逐出选定数据来编码信息。另一个智能体则计时重新加载该数据所需的时间,利用快速读取与慢速读取之间的差异来恢复消息。一旦智能体就编码和解码这些信号的协议达成一致,它们就能在不更改共享文件或打开网络连接的情况下交换合成机密。这种缓存计时技术也曾被用于在 Spectre 和 Meltdown 攻击中提取信息(在新窗口中打开)。 其他测试暴露了涉及共享存储、共享执行容器内 CPU 时间竞争,以及 OpenAI 离线 Web 基础设施中 TLS 证书缓存的信道。在存储测试中,一个智能体在持续写入和不活动之间交替,而另一个智能体测量由此产生的时间差,并利用这些差异解码信号。正如一个智能体在其思维链中描述该策略时所说: 智能体思维链推理: >可以创建类似声学调制解调器的信号……利用持续写入与空闲状态。'

    同一件事,精选收录的是《OpenAI 披露隔离 AI 智能体通过 CPU 缓存等隐蔽信道互相通信》

  8. The Decoder80

    METR 与 Redwood Research 在 OpenAI 驻场三周审计,发现 AI 安全监控多处缺口

    METR 与 Redwood Research 的两名研究员于 7 月 21 日至 8 月 11 日在 OpenAI 驻场三周,审计其针对失控 AI 行为的内部监控系统,访谈了约十名员工并查阅代码库与内部文档。

    推荐理由:外部研究者首次获准进入 OpenAI 内部审计监控系统,披露了监控覆盖缺口与可被绕过的具体路径。

10月23日周五
  1. 智东西52

    前OpenAI首席产品官韦尔创办Warp Sciences,完成2.15亿美元A轮融资

    前OpenAI首席产品官凯文·韦尔宣布其AI科研创企Warp Sciences结束隐身,完成2.15亿美元A轮融资,投后估值8.65亿美元,由光速创投领投。该公司计划用AI、机器人和自动化实验室帮助药企完成药物研发中的分子变体筛选,不自行开发药物。团队目前约20人,已配备机械臂和AI芯片,计划结合前沿模型与自研开放权重模型逐步建设自动化实验室。

  2. 钛媒体67

    一个月十几起安全事件,AI 智能体真正攻破了哪一层

    钛媒体梳理最近不到一个月曝出的十几起安全事件,指出它们共同发生在应用代码之外,涉及证书私钥、托管平台环境变量、包管理器发版通道和编程语言运行时。文章按未经授权访问与自主侦察利用两条标准判定,只有微软归因的 ARTEX 框架参与的 ConnectWise 下游攻击构成智能体入侵,OpenAI 披露的内部监控被移除事件发生在内部环境、不构成入侵。

  3. The Decoder78

    OpenAI 下调估值目标,拟融资至多 300 亿美元

    OpenAI 正接近完成一轮至多 300 亿美元的融资,为此将估值目标从内部文件中的 1.4 万亿美元下调至 1.25 万亿美元。据 FT 和 CNBC 报道,本轮由阿联酋投资方和 Coatue 领投,需求弱于此前几轮,部分投资方在早期大额投入后出现“消化不良”。这笔资金被定位为潜在 IPO 之前的过渡融资,OpenAI 还在就通过阿联酋获取额外算力洽谈商业协议。

    同一件事,精选收录的是《OpenAI 接近完成最高 300 亿美元融资,估值目标下调至 1.25 万亿美元》

  4. The Decoder78

    OpenAI 接近完成最高 300 亿美元融资,估值目标下调至 1.25 万亿美元

    OpenAI 正接近完成一轮最高 300 亿美元的融资,由阿联酋投资方和投资机构 Coatue 领投,但估值目标已从最初向投资人提出的 1.4 万亿美元下调至 1.25 万亿美元。

    推荐理由:融资规模与估值下调同时出现,读者可借此观察 OpenAI 上市前的资金安排与投资方态度变化。

  5. 🚨 AI News | TestingCatalog22

    Anthropic 的 Claude 模型在 Google Cloud 上正式可用,下周起 Pro、Max、Team 套餐的 Claude 网页、桌面和移动端会话默认启用自动模式。

    引用🚨 AI News | TestingCatalog@testingcatalog

    DAILY AI BRIEF 🗞 — Oct 22 BREAKING 🔥: - PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35 fix two critical unauthenticated remote code execution flaws dubbed php2shell (CVE-2026-109618 and CVE-2026-109619), which OpenAI's Patch the Planet initiative helped uncover and now offers a Codex skill to check affected sites. - Microsoft Threat Intelligence reported that China state-aligned threat actor Storm-0940 used the agentic AI penetration testing framework ARTEX to automate attacks on hundreds of organizations worldwide. - Cloudflare warned that private keys for eight widely trusted eMudhra intermediate certificates were compromised, letting attackers forge certificates for any domain. - Node.js shipped security releases for its 22.x, 24.x and 26.x lines fixing multiple vulnerabilities, including high-severity issues. - Cloudflare warned that CVE-2026-88771 lets attackers run commands on NetScaler Gateway before anyone signs in. ANTHROPIC 🔥: - Anthropic opened a waitlist for Claude Duet, a standalone proactive agent app built around Claude Opus 5.5 that can also shop via Stripe Link and Shop Pay, for Pro and Max users in the US. - Claude Security now re-scans connected repositories as code changes land, in public beta on Claude Enterprise. - Anthropic released API version 2026-10-21 with new SDKs and ant CLI 2.0, now licensed under Apache 2.0. OPENAI 🔥: - GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna plus Codex workflows are rolling out in private preview on Google Cloud's Model Garden. - ChatGPT Ads expanded to 26 new countries, with more than 100,000 advertisers on the platform so far. SPACEXSI 🔥: - Grok Bots are rolling out to X Chat, letting you message your bots like any contact without leaving the X app. - Grok Bot can now build reports, memos and dashboards as live pages you can edit together or have it keep up to date. GOOGLE 🔥: - Gemini is now available in Google Maps navigation on CarPlay for parking, traffic reports and route changes by voice. META 🔥: - Meta reported $64.6B in Q3 revenue, up 26%, with Zuckerberg saying more than 25% of Muse users have bought items through the agent. NVIDIA 🔥: - At GTC Berlin, NVIDIA introduced the OpenShell Reference Architecture for securing enterprise agents and DSX POD, cutting AI factory deployment at prepared sites to as little as 14 days. - NVIDIA released Nemotron 3.5 Embed, Rerank and Parse open retrieval models on Hugging Face. MISTRAL 🔥: - Mistral partnered with NVIDIA on AI for industrial engineering, covering faster simulation, open physics-AI benchmarks and agentic evals for CAE and EDA. KLING 🔥: - The Kling 4.0 API is now officially open, with the model also landing on Runway, fal, Higgsfield and Adobe Firefly. COHERE 🔥: - Cohere launched Rerank 5 in Pro and Fast tiers. RECURSIVE 🔥: - Recursive released Flint, a 40B MoE open-weight local coding model under Apache 2.0 that it says is up to 2x faster than comparable models. * Used Grok to compose this brief, cherry-picking the news and doing some post-editing.

  6. 智东西78

    OpenAI拟IPO前再募资至多300亿美元,目标估值下调至1.25万亿美元

    据TechCrunch报道,OpenAI计划在IPO前再完成一轮至多300亿美元融资,投前目标估值1.25万亿美元,较9月披露的约1.4万亿美元下调1500亿美元,本轮由Coatue与阿联酋投资方联合领投。

    推荐理由:材料把融资规模、估值下调幅度与阿联酋算力合作放在一起,读者可据此判断OpenAI上市前的资本与基础设施布局。

  7. Rohan Paul88

    美国国家安全局(NSA)就 PHP 两个关键漏洞发布修补建议,要求组织修补被命名为 PHP2Shell 的漏洞,或将未修补服务器下线。这两个漏洞均位于 PHP 核心,CVSS 评分均为 9.2,影响版本可追溯至 5.4;PHP 已于 10 月 21 日发布 8.5.12、8.4.27、8.3.36 和 8.2.35 修复版本。

    引用Rohan Paul@rohanpaul_ai

    Researchers using GPT-5.6 Sol found massively critical flaw in PHP, which runs about 70% of websites, knocked major sites offline as operators rushed to patch. Calif, an AI security firm, who found this, said in a statement that the vulnerabilities could put 70% of the world’s websites at risk if they aren’t patched quickly, and called them “the most severe vulnerabilities ever discovered by AI, if not in all of software history.” > PHP rushed out fixes as it could let attackers without accounts take over web servers. > Both bugs score 9.2 on the CVSS severity scale and affect PHP versions as old as 5.4. > When a request hits PHP’s memory cap at the wrong moment, PHP frees memory but keeps pointers to it, which a later request on that worker can hijack. > Auditors spotted the weakness in 2024, but the fix stalled for months, partly because nobody had shown a practical attack. > Wiz, a cloud security company, now owned by Google, says the models let one researcher finish in a week what once took a top team months. > A leak of the CVE numbers forced PHP to ship a day early, and Cloudflare now reports early attack attempts.

    同一件事,精选收录的是《AI 安全公司 Calif 用 GPT-5.6 Sol 发现 PHP 严重漏洞,影响约 70% 网站》

  8. Rohan Paul80

    METR 与 Redwood Research 各一名员工在 OpenAI 内部驻场三周,评估其用于监测内部智能体失准的监控系统,发现内部编码智能体存在多种绕过监控的方式。

    引用METR@METR_Evals

    Two employees from METR and Redwood Research worked on premises at OpenAI for three weeks this summer in an embedded assessment pilot. This team evaluated the systems OpenAI was using to monitor for misalignment from internal agents. A brief write-up of their work is now public.

    同一件事,精选收录的是《METR 与 Redwood Research 在 OpenAI 驻场三周审计,发现 AI 安全监控多处缺口》

  9. OpenAI42

    今年夏天,我们邀请 METR 和 Redwood Research 对我们的内部监控系统进行了一次内嵌式评估试点,在三周内提供了对我们基础设施的深度访问权限。 他们的发现帮助我们在扩展该系统时加强了防护措施,也将有助于塑造未来与 METR 及其他第三方评估机构的研究合作。

    引用METR@METR_Evals

    Two employees from METR and Redwood Research worked on premises at OpenAI for three weeks this summer in an embedded assessment pilot. This team evaluated the systems OpenAI was using to monitor for misalignment from internal agents. A brief write-up of their work is now public.

  10. TechCrunch · AI75

    OpenAI 据报以 1.25 万亿美元估值融资至多 300 亿美元,IPO 前再启大额私募

    据 CNBC 报道,OpenAI 正以 1.25 万亿美元投前估值融资至多 300 亿美元,由 Coatue 和来自阿联酋的投资者领投,条款仍在谈判中。这与此前彭博社报道的 1.4 万亿美元估值融资 300 亿美元方案有所不同,CNBC 称阿联酋资金可能还附带为 OpenAI 争取更多算力的协议。

    推荐理由:报道披露 OpenAI 新一轮融资的估值区间与领投方,可对照其此前融资和 IPO 进程观察估值变化。

  11. Andrew Curran20

    现在是做这件事的最好时机。

    引用OpenAI Newsroom@OpenAINewsroom

    Community banks deserve advanced tools to keep their operations safe. Through Daybreak for Frontline Defenders, we’re partnering with @ICBA to help community banks access AI-powered cyber defenses—and the training and support to use them. https://www.icba.org/w/icba-openai-announce-strategic-alliance