#网络安全

indigo
6天前
Capybara 是 Claude 下一代全新旗舰层级,Mythos 是它的代号,能力比 Opus 4.6 有阶跃式提升,网络安全能力尤其突出到让 Anthropic 自己都担心——是目前已知最接近 AGI 级别安全风险的商业模型。 3月26日,Anthropic 因 CMS(内容管理系统)配置错误,意外将约 3000 个未发布资产(包括一篇草稿博客)暴露在公开可搜索的数据缓存中。两位网络安全研究员(LayerX Security 的 Roy Paz + 剑桥大学的 Alexandre Pauwels)独立发现,Fortune 独家报道。Anthropic 被通知后迅速关闭了访问权限。 关于 Capybara / Mythos: - Capybara 是 Anthropic 创建的全新模型层级,比现有 Opus 系列更大、更强、更贵——也就是说,未来的产品线将是 Haiku / Sonnet / Opus / Capybara(四级); - Claude Mythos 是基于 Capybara 训练出来的具体模型,草稿博客称其为「迄今为止我们开发的最强 AI 模型,没有之一」; - 对比 Opus 4.6,Capybara 在代码、学术推理、网络安全等测试中「得分大幅提升」; - Anthropic 确认:这是「能力上的阶跃式变化」 最大的发布顾虑:网络安全风险!草稿博客中 Anthropic 自己写道: - 该模型「在网络安全能力上目前远超任何其他 AI 模型」 - 它「预示着即将到来的一波 AI 模型,其漏洞利用能力将远超防御者的努力」 - 因此发布策略极为谨慎:优先向网络安全防御机构提供早期访问权限,让防御者提前「给代码库打疫苗」
NVIDIA NemoClaw:NVIDIA 为 OpenClaw 打造的安全插件,在本地搭建沙箱环境,让 Agent 24*7 自主运行的同时,用策略管控其网络、文件、进程和推理调用。 NemoClaw 有什么用? OpenClaw 正在成为个人 AI 的"操作系统",一个始终在线、能自我进化的自主助手。但自主性越强,安全风险越大:Agent 可能访问不该访问的网络、读写敏感文件、调用未授权的模型。NemoClaw 要解决的核心问题是:不牺牲自主性,给 Agent 加上可控的安全边界。 做法是引入 NVIDIA OpenShell 运行时,用声明式策略管控 Agent 的一切外部交互。 架构四件套 · Plugin:TypeScript CLI,用户直接操作的入口——启动、连接、查状态、看日志 · Blueprint:版本化的 Python 制品,编排沙箱创建、策略应用和推理配置。生命周期四步走:解析制品 → 校验摘要 → 规划资源 → 应用部署,借鉴了 IaC 的思路 · Sandbox:受策略约束的 OpenShell 容器,OpenClaw 的实际运行环境 · Inference:所有模型调用经 OpenShell 网关中转,Agent 本身无法直接外连 四层防护 · 网络层:阻断未授权出站连接,运行时可热加载白名单 · 文件系统层:读写范围锁定在 /sandbox 和 /tmp,沙箱创建时固化 · 进程层:阻止提权和危险系统调用,沙箱创建时固化 · 推理层:模型 API 调用重定向到受控后端,运行时可热加载 底层隔离技术栈:Landlock(文件访问控制)+ seccomp(系统调用过滤)+ network namespace(网络隔离),比普通容器隔离更细粒度。 一个关键设计:Agent 尝试访问白名单外的主机时,OpenShell 拦截请求并在 TUI 中弹出审批,实现"人在回路"的安全模式。 推理路由 Agent 的推理请求全部被网关拦截后路由,当前默认走 NVIDIA 云端的 nemotron-3-super-120b-a12b 模型,需要 API key。本地推理(Ollama、vLLM)还处于实验阶段。 这个设计既做推理审计和控制,也为未来的"隐私路由器"铺路——按数据敏感度自动分流本地或云端模型。
The leaked files from Chinese company 中科天璣 (GoLaxy/Zhongke Tianji), a spin-off from the Chinese Academy of Sciences, expose a massive, AI-powered surveillance and infiltration operation targeting Taiwan. This 399-page internal document set, first surfaced in August 2025 via Vanderbilt University's researchers and covered by The New York Times, was deeply analyzed for the Taiwan context by Taiwan's Doublethink Lab (台灣民主實驗室). Key revelations include: - Full profiling of Taiwan's 23 million population via 22 million closed-source household registration data. - Detailed "涉台政要图谱" (Taiwan political elites map) dossiers on leaders like President Lai Ching-te (賴清德), former President Tsai Ing-wen (蔡英文), Ko Wen-je (柯文哲), and others across parties. - Profiles include party affiliation, position, education, religion, social connections, and crucially a "对华态度" field labeling them as pro-China / anti-China, categorized into stubborn, friendly, swing, and objective factions (each with at least 1000 entries) clearly designed for divide-and-conquer tactics. - Massive data haul: 6million news and social media items, at least 10,000 global Taiwan-related tracking sources, ~5000 key social accounts targeted, plus mappings of 75 political parties, 1478 companies, 13,000+ religious groups, and nearly 24,000 civic organizations — creating a near-complete panorama of Taiwan's civil society. - AI-driven deception: Use of "identity shaping technology" to generate hyper-realistic virtual personas fluent in Mandarin, Taiwanese Hokkien (台语), and English. These fake identities impersonate real Taiwanese people online for cognitive warfare, social division, and influence ops. At least 7 such operated fake accounts already identified. This isn't isolated to Taiwan — the same toolkit has been applied in Hong Kong (monitoring 180,000 people at one point), Xinjiang, the US, and likely beyond. The system combines big data scraping, AI profiling, automated influence, and bot networks to undermine democratic societies from within. Taiwan's democracy faces sophisticated, state-linked digital threats. This leak is a wake-up call for awareness, resilience, and international scrutiny of PRC-linked tech firms. Share widely to expose these operations — transparency defeats covert control. Sources: Doublethink Lab analysis, Liberty Times (自由時報) reporting, original GoLaxy leaks via Vanderbilt/ NYT coverage. #Taiwan #中科天璣 #ChinaSurveillance #AIFakeAccounts #CognitiveWarfare #PRCInfluence