←搜索结果

按匹配度 · 共 152 条结果 清除搜索

星期二

  1. Replit Blog

    Replit 上线黑盒渗透测试扫描

    Replit 推出黑盒渗透测试扫描,以外部攻击者视角通过网络和浏览器测试应用,无需访问源码,与已有的白盒代码扫描互补。扫描在私有沙箱中针对应用副本运行,先以未登录访客身份探测,再以普通登录用户身份检查越权访问,最后给出可确认的问题清单供 Replit Agent 修复。Replit 表示两种扫描的发现很少重叠,例如白盒扫描发现权限撤销后仍可操作,黑盒扫描发现管理后台地址可猜且无需登录。该功能加入 Replit Auto-Protect,平台现有 3 级临时安全扫描,Level 3 同时运行白盒与黑盒扫描。

    Black-box pen tests on Replit — cover image

    ng a pen test from a vendor that would cost thousands, and weeks of back-and-forth. A penetration (”pen”) test is a safe, authorized mock cyberattack that helps you fix security flaws before bad hackers find them. As we have democratized software creation, we have also sought to democratize securing it. A new generation of creators needs to ensure their apps are hardened against attackers, who look at apps built with AI as an opportunity as a potential opportunity. Our existing scans have an age

    匹配字段正文

星期六

  1. Augment Code 博客

    Augment Code 重建 Auggie CLI 架构,任务成本比 Claude Code 低 53%

    Augment Code 宣布重建 Auggie CLI 的 harness,在 SWE-bench Pro 上相同通过率下,Auggie v2 完成任务成本为 $1.27,比 Claude Code 的 $2.70 低 53%。重建基于开源项目 Pi,通过精简工具面、用检索代替探索和主动压缩上下文实现效率提升,CLI 将于 9 月或 10 月推出。

    How we rebuilt the Auggie CLI harness to make tasks 53% cheaper than Claude Code

    inal, so every rendering bug became a debugging session through a React reconciler and a flexbox engine. And there were limits we never engineered around: the whole-transcript redraws flickered, and layout occasionally came out wrong. The AI SDK and Mastra prototypes moved to OpenTUI, which maintains a virtual terminal buffer and keeps it in sync with the real one. That fixes flicker, but you give up the native buffer along the way: the transcript is no longer plain text in your terminal's scrol

    匹配字段正文

星期五

  1. The Register

    ShieldFont 开源字体用同词性替换向 AI 爬虫投毒

    阿姆斯特丹设计工作室 Seneda & Abrucio 与哥本哈根字体公司 Playtype 合作推出开源项目 ShieldFont,通过扩展 OpenType 的 GSUB 字形替换表,让网页在人类眼中正常显示、而爬虫读取的原始 HTML 中约四分之一词汇被替换成同词性同语义类别的其他词,从而向训练数据注入难以分辨的噪声。项目已提供在线编码器、React 组件以及 CSS 和 CDN 集成方式,桌面字体约 5 MB、压缩网页字体约 800 KB,并附带三套 GSUB 字典且支持自建字典以防逆向。团队承认它可被 OCR 截图、下载字体跑字典等方式绕过,也可能影响 SEO、翻译应用、系统复制粘贴和屏幕阅读器,目前处于 v0/alpha 阶段。

    tonking Stingray' beta swims out: Deeper Rust 'oxidization' plus LLM speech-to-text GNOME 51 is here, and Ubuntu interim releases are getting interesting again; Revive Raskin's 3 laws of software with humane work that considers a dev's duty of care All this, and angry NeoVim users; Firefox 157: Could bold new look be just what Mozilla needs? Yet another new version of Firefox, with a significant reskin; KDE turns 30 and someone's brought an AI-native desktop proposal Akademy talk imagines Plasma

    匹配字段正文

星期四

  1. Replit Blog精选

    Replit 发布 Replit Design 设计套件

    Replit 发布 Replit Design 创意设计套件,面向所有用户开放,可从想法直接生成设计并发布,无需导出或重建。该套件支持 Claude、GPT-5、Gemini、Kimi、GLM 等模型,内置 Ambient Intelligence 逐步给出可一键采纳的变体建议,并集成 Mobbin 的 60 万张真实界面素材与 1000 多个应用参考。Replit Design 是 Canvas 的下一代演进,原有 Canvas 项目全部保留,官方同时启动奖金超过 5 万美元的 Designathon 活动。

    Introducing Replit Design — cover image

    t now. And in an industry first, everyone gets access. No Mobbin account required; it’s just there, wired in. To use it, just describe what you want to create, and Replit Design draws from more than 600,000 real-world screens from over 1,000 apps to turn your idea into a finished app or site. Your brand, everywhere, in one click AI design today is fragmented. Prompt in one tool, refine in another, publish in a third. Every handoff strips something out: the colors drift, the type resets, the deta

    匹配字段正文

星期三

  1. The Register

    Perplexity 将 Model Council 扩展至 Computer 平台,最多可选 8 个模型

    Perplexity 周二把上线五个月的 Model Council 扩展到云端 Computer 平台,用户可选 2 到 8 个模型组成顾问团,模型各自独立作答后由用户选定的模型担任主席,综合出报告并标出结论一致与分歧之处。该功能此前仅限企业版和 Max 用户,现在也向个人 Max 和 Pro(20 美元/月)用户开放,可选模型涵盖 OpenAI、Anthropic、Google 以及 GLM、Kimi 等开放权重模型。Computer 采用按量计费,100 credits 等于 1 美元,Pro 用户获 4000 赠送 credits,Max 用户获 35k 及每月续期的 10k。

    tonking Stingray' beta swims out: Deeper Rust 'oxidization' plus LLM speech-to-text GNOME 51 is here, and Ubuntu interim releases are getting interesting again; Revive Raskin's 3 laws of software with humane work that considers a dev's duty of care All this, and angry NeoVim users; Firefox 157: Could bold new look be just what Mozilla needs? Yet another new version of Firefox, with a significant reskin; KDE turns 30 and someone's brought an AI-native desktop proposal Akademy talk imagines Plasma

    匹配字段正文

星期一

  1. LMSYS Blog精选

    SGLang 与 Miles 为 Kimi K3 提供 Day-0 支持

    SGLang 与 Miles 宣布在 Kimi K3 发布当天提供 Day-0 支持,前者负责推理、后者负责 RL 训练,与 Moonshot AI、NVIDIA 团队合作完成。K3 是首个 3 万亿参数级别的开源模型,实际为 2.8T 参数、1M token 上下文窗口并具备原生视觉理解,其 69 层 KDA 线性注意力与 24 层 MLA 交织的混合架构打破了常规服务栈的多项假设。文中给出统一显存池、DSpark 投机解码与 ReplaySSM、分阶段并行等方案,并报告在 16 节点 × 4 GB300 上做 DAPO 数学 RL 训练,AIME-2024 贪心评测在 12 小时内从 43.3% 升至 76.7%。

    The three KDA state moves, copy-on-write, snapshot and donate, and where each sits relative to the serial forward stream.

    s flat at the 2e-3 MXFP4 floor for the whole run. The 12-hour DAPO run. Left: mean reward over the 64 sampled responses per rollout, graded on the response channel; each rollout draws a different prompt batch, so the trend is the signal and the per-step value is noise. Middle: AIME-2024 greedy pass rate, evaluated every 10 rollouts under the same 4096-token limit as training, so a correct solution that runs past the limit scores as a miss. Right: a Schulman k3 estimate of KL(rollout ‖ train) at

    匹配字段正文

星期四

  1. Black Forest Labs Blog

    Black Forest Labs 与 mimic 推出基于 FLUX 3 的视频-动作模型 FLUX-mimic

    Black Forest Labs 发布 FLUX 3 早期版本,并与机器人公司 mimic 合作推出视频-动作模型 FLUX-mimic,已在 Audi 测试和部署。FLUX 3 从训练之初就联合训练图像、视频和音频,视频预测占总算力成本 95% 以上;加入动作预测后,文生视频和图生视频的人类评分一度下降最多 10%,3500 步后恢复原有质量并同时预测动作。FLUX-mimic 在冻结 FLUX 主干的情况下,动作解码器表现超过此前的视觉-语言-动作模型,微调后达到 SOTA 成功率;主干在单张 NVIDIA RTX 5090 上从输入到世界表征耗时不到 80ms,整套机器人系统反应时间为 101ms。

    FLUX 3 x mimic: The Next Generation of Video-Action Models

    ity have long been studied and approached in isolation from each other. Generative approaches result in high-quality world models that enable simulations and they exhibit scaling laws for predictable returns on compute investments. However, compared to more specialized approaches for representation learning they produce less disentangled representations, which puts a ceiling on their usefulness for tasks that require world understanding. As generative models themselves rely on their own features

    匹配字段正文

星期三

  1. Hugging Face

    一键从 Hugging Face 进入 Amazon SageMaker Studio

    Hugging Face 为受支持模型新增直达 Amazon SageMaker Studio 的按钮,可预载模型并进入微调或端点部署流程。新建 Studio 环境会预配置训练、模型定制和部署权限,界面还会直接显示 G5、G6 GPU 实例的配额可用性。

    d reinforcement learning from AI feedback (RLAIF), with supported deployment to SageMaker AI or Amazon Bedrock endpoints. This alleviates the need to manually create and configure AWS Identity and Access Management (IAM) roles and policies before you can start experimenting. For existing Studio environments, actionable messages with direct links to documentation guide you through adding these permissions. GPU quota visibility When selecting instance types for deployment or training, the Studio U

    匹配字段正文

星期二

  1. Hugging Face精选

    使用 SkyPilot 在任意云运行 AI 工作负载,并以零读取流量费存储于 Hugging Face

    Hugging Face Storage 已成为 SkyPilot 的一级存储后端,任务可通过 hf:// 在不同云端挂载模型、数据集和 Bucket,并免收读取流量费。其 FUSE 挂载支持懒加载和本地缓存,测试中 Qwen/Qwen3.5-4B 约 30 秒即可开始训练;写回仍会产生计算云通常收取的流量费,MOUNT 还要求 glibc 2.34+ 和 /dev/fuse。

    HF TOKEN in your environment and hand it to a run with --secret HF TOKEN ; SkyPilot uses it for the mount on whatever cloud the job lands. One token works whether the job lands on AWS, GCP, Azure, Nebius, Lambda, or your own Kubernetes cluster, so there are no per-cloud bucket keys to juggle. No egress: storage stops deciding where you run GPU capacity rarely comes from one place anymore. To get enough H100s and H200s, teams hold reserved and committed capacity across several vendors at once (a

    匹配字段正文

星期四

  1. Cognition 模型 / Devin 博客精选

    Cognition 发布 Devin Security Swarm,自动验证漏洞并提交修复 PR

    Cognition 发布 Devin Security Swarm,由并行智能体分段排查代码库,组合出完整攻击路径并在隔离沙箱中复现验证可利用性,确认后自动编写补丁并提交 PR 供审查。官方称其在 50 个真实漏洞基准上比最接近的同类方案多发现漏洞且成本低 30%,并独有发现 PHP 模板注入沙箱绕过等三个关键漏洞。该产品即日起可用,支持从威胁模型文档生成扫描配置、按日/周/自定义计划扫描,首次全量扫描后仅处理变更代码。

    st than the nearest comparable alternative. How it works A swarm of parallel agents investigate segments of the codebase. Each agent reasons across files—catching business logic flaws, chained auth bypasses, and cross-service exploit paths. Devin then composes individual findings into full attack paths and reproduces each one in an isolated sandbox to confirm exploitability at runtime. What reaches the security team is a set of confirmed-exploitable vulnerabilities with attack paths and steps to

    匹配字段正文

星期三

  1. The Register

    Anthropic 推出 Claude Tag,取代 Slack 版 Claude 应用

    Anthropic 将停用现有 Claude in Slack 应用,改为推出常驻智能体 Claude Tag,它可以作为团队成员加入 Slack 频道,访问管理员指定的频道、工具、数据和代码库,用户 @Claude 即可委派任务。Anthropic 称其产品团队 65% 的代码由 Tag 编写,Tag 还能跨频道学习、主动推送更新并自行安排任务。现有连接器将于 8 月 3 日停止服务,Enterprise 和 Teams 客户即日起可用 Tag,迁移窗口为 30 天,迁移的 Enterprise 客户可获 25000 美元额度、Teams 客户 2500 美元,额度 9 月 1 日到期。

    tonking Stingray' beta swims out: Deeper Rust 'oxidization' plus LLM speech-to-text GNOME 51 is here, and Ubuntu interim releases are getting interesting again; Revive Raskin's 3 laws of software with humane work that considers a dev's duty of care All this, and angry NeoVim users; Firefox 157: Could bold new look be just what Mozilla needs? Yet another new version of Firefox, with a significant reskin; KDE turns 30 and someone's brought an AI-native desktop proposal Akademy talk imagines Plasma

    匹配字段正文

  2. DevClass

    Vercel 发布开源智能体框架 eve 并推出 Passport

    Vercel 在伦敦 Ship 活动上发布开源智能体框架 eve,用 TypeScript 和 Markdown 定义指令、技能、模型提供方、工具、认证、渠道和调度,智能体默认运行在隔离 VM 沙箱中,代码以 Apache 2.0 许可托管在 GitHub。eve 构建的智能体默认通过 vercel deploy 部署到 Vercel,也可用 npx eve dev 在本地运行,模型可接入 AI SDK 支持的任何模型。Vercel 同时预览 Enterprise Apps and Agents,其中 Passport 基于 OpenID Connect 把团队内应用和 AI 智能体统一置于 Okta、Microsoft Entra 等身份提供商之下,用于应对员工用 AI 生成、却脱离企业 IT 管控的 shadow AI 应用。

    ctory sync to enable Vercel in a team to be managed by the organization's identity system. Finally, Bring Your Own Cloud (BYOC) lets organizations use Vercel's platform running on AWS infrastructure provisioned by the customer. According to Vercel, Passport was a highly requested feature because of the number of employees who create applications hosted by Vercel but outside the control of the organization. A typical scenario is that an employee builds an application with AI assistance, and the A

    匹配字段正文

星期四

  1. Claude Blog

    Claude Code 支持 artifacts,可将会话工作转为可共享的实时页面

    Anthropic 宣布 Claude Code 开始支持 artifacts,能把会话中的工作进度转成可共享的实时可视化页面,例如 PR 讲解、系统说明、仪表盘和发布清单,并随会话进展自动更新。artifacts 基于会话的完整上下文构建,包括代码库、连接器和对话本身,无需自行接入数据源或搭建基础设施;每次发布在同一链接生成新版本并保留版本历史,页面默认仅作者可见,可共享给团队成员和组织,且不能设为公开。该功能以 beta 形式面向 Claude Team 和 Enterprise 组织开放,可在 Claude Code CLI 和桌面应用中使用,页面可在任意浏览器查看。

    er blog Console Console Console Engineering at Anthropic Engineering at Anthropic Engineering at Anthropic Overview Overview Overview Marketplace Marketplace Marketplace Claude on AWS Claude on AWS Claude on AWS Blog Blog Blog Claude partner network Claude partner network Claude partner network Customer stories Customer stories Customer stories Powered by Claude Powered by Claude Powered by Claude Service partners Service partners Service partners Availability Availability Availability Check fil

    匹配字段正文

星期三

  1. xAI News精选

    xAI 的 Grok 4.3 在 Amazon Bedrock 正式可用

    xAI 宣布 Grok 4.3 在 Amazon Bedrock 正式可用,开发者可通过 Bedrock 的推理引擎调用该模型。Grok 4.3 支持 100 万 token 上下文窗口,推理强度可配置为 none、low、medium、high,定价为输入 $1.25 / 百万 token、输出 $2.50 / 百万 token。xAI 称其在 Artificial Analysis Omniscience 基准上取得最低幻觉率,并在 Tau2 Telecom、Vals AI Case Law 与 Corporate Finance 基准上排名第一。

    lligence per dollar than other frontier models. Input : $1.25 per 1 million tokens Output : $2.50 per 1 million tokens It features a 1 million context window that handles extremely long documents and codebases with ease. Developers can also configure reasoning: Use none for maximum speed on simple requests, or low / medium / high to scale reasoning effort to the difficulty of the problem. Build with Grok 4.3 on Amazon Bedrock Grok 4.3 is now available to all developers in supported AWS Regions .

    匹配字段正文

星期五

  1. Linear:Now精选

    Linear 推出编码会话、自动分诊和 Diffs,让 Linear Agent 直接编写代码

    Linear 推出编码会话、自动分诊和 Diffs,让 Linear Agent 能从 issue 出发调查代码库、编写代码并创建 pull request。任务在云端使用 Claude Code 或 Codex 等模型与工具执行,团队成员可共同跟进、补充上下文和审查变更。Linear 称其上月已有近 700 个由该智能体完成的 PR 被合并。

    Abstract image of circles forming a loop.

    ssue the moment it’s triaged rather than waiting to be told what to do. You can set up the agent to first investigate the issue. It reads the codebase with Code Intelligence and draws on the context around the issue, like the discussions and affected customers. The agent can also connect to your team’s other tools like Sentry or Datadog to triangulate the problem. With a grasp of what to do, and the confidence there is a fix, the agent can open a coding session in the cloud, write a fix, open a

    匹配字段正文

星期一

  1. Claude Blog

    Anthropic 发布 Swift 包,让 Apple 开发者通过 Foundation Models 框架调用 Claude

    Anthropic 发布新的 Swift 包,为 Apple 的 Foundation Models 框架提供 Claude 支持,开发者可在本地模型处理摘要、抽取等任务后,把需要多步推理、代码生成或联网搜索的请求交给 Claude。该包处理流式输出、工具调用和结构化响应,可直接回填到 SwiftUI 视图,需使用 Anthropic API key,支持 iOS 27、iPadOS 27、macOS 27、visionOS 27 和 watchOS 27。

    er blog Console Console Console Engineering at Anthropic Engineering at Anthropic Engineering at Anthropic Overview Overview Overview Marketplace Marketplace Marketplace Claude on AWS Claude on AWS Claude on AWS Blog Blog Blog Claude partner network Claude partner network Claude partner network Customer stories Customer stories Customer stories Powered by Claude Powered by Claude Powered by Claude Service partners Service partners Service partners Availability Availability Availability Check fil

    匹配字段正文

星期四

  1. The Register

    微软发布 Autopilot 智能体类别,首个智能体 Scout 亮相

    微软在 Build 大会上发布名为 Autopilot 的新一类智能体,首个智能体 Scout 可在 Teams、Outlook、OneDrive 和 SharePoint 间运行,连接聊天、邮件、日历和联系人数据,并能自主安排会议、标记重要会议、生成准备材料、识别临近截止时间并预留日程。Scout 由 OpenClaw 驱动,绑定 Entra 身份并受组织访问控制约束,目前仅向少数客户和 Frontier 计划组织开放预览。微软称其具备企业级安全与管控,但文章指出 OpenClaw 的安全口碑存疑,且未说明针对常见 AI 攻击的其他防护措施。

    tonking Stingray' beta swims out: Deeper Rust 'oxidization' plus LLM speech-to-text GNOME 51 is here, and Ubuntu interim releases are getting interesting again; Revive Raskin's 3 laws of software with humane work that considers a dev's duty of care All this, and angry NeoVim users; Firefox 157: Could bold new look be just what Mozilla needs? Yet another new version of Firefox, with a significant reskin; KDE turns 30 and someone's brought an AI-native desktop proposal Akademy talk imagines Plasma

    匹配字段正文

星期三

  1. MiniMax Blog

    MiniMax 发布 Agent Team,主智能体拆分任务并内置对抗式验证

    MiniMax 将 Agent 整体升级并命名为 Mavis,同时推出 Agent Team:由主 Agent 把复杂任务拆成并行子任务,交给一批 Agent 执行,并内置对抗式质量门。团队简化为 Leader、Worker、Verifier 三类角色,Worker 与 Verifier 形成对抗关系,验证失败时 Team Engine 会重新唤醒生产节点继续修改。官方称最新版 MiniMax Agent 即将开源,预计与 MiniMax M3 模型一同发布,桌面端应用已上线。

    MiniMax Agent Team Banner

    logs, file artifacts, and decision records as recoverable objects. Agent collaboration is a stateful long-term task. 3.2. Core scenario 2: Coding Harness The Agent Team project draws heavily on the Harness mindset. Harness pushes a step further from “Agent writes code”: an Agent must follow the full development lifecycle — code lives on branches, execution in sandboxes, edits as diffs, tests that can rerun, reviews that leave records, failures that can be replayed, and when needed, tasks that c

    匹配字段正文

  2. Augment Code 博客

    Augment Code 用 Cosmos 智能体将事故响应人工调查工作量降低 81%

    Augment Code 发布博客,介绍其 Cosmos 平台中的 Incident Investigator 智能体如何在 Slack 内自动执行事故调查与根因分析。部署后,智能体处理了 81.3% 的事故,人工调查工作量减少约 81%,首次 RCA 中位时间从 30.1 分钟降至 6.2 分钟,MTTR 从 29.5 分钟降至 19.9 分钟。

    Scaling incident management for an AI-native organization using Cosmos

    of the most time-consuming parts of the SDLC. Before Augment, Akshay spent several years at Uber advancing automated code review and repair systems, and conducted research across AWS’s Automated Reasoning Group, Google’s Android Static Analysis team, and UCLA. His work sits at the intersection of AI, software engineering, and programming-language theory. Sophie Reynolds Sophie Reynolds is a Software Engineer at Augment Code. She joined full-time in early 2025 after a Codepoint Fellowship with S

    匹配字段正文

星期二

  1. Claude Blog精选

    Claude Managed Agents 新增自托管沙箱与 MCP 隧道

    Claude Managed Agents 现在可以在用户自控的沙箱中执行工具,并连接私有 MCP 服务器,沙箱与所访问服务都运行在企业边界内。自托管沙箱在 Claude Platform 上处于公开测试,支持 Cloudflare、Daytona、Modal、Vercel 等托管沙箱提供商;MCP tunnels 处于研究预览阶段,需申请访问,可在 Managed Agents 和 Messages API 中使用,由组织管理员在 Claude Console 的工作区设置中管理。智能体的编排、上下文管理和错误恢复仍运行在 Anthropic 基础设施上,工具执行则移到用户配置的环境中。

    er blog Console Console Console Engineering at Anthropic Engineering at Anthropic Engineering at Anthropic Overview Overview Overview Marketplace Marketplace Marketplace Claude on AWS Claude on AWS Claude on AWS Blog Blog Blog Claude partner network Claude partner network Claude partner network Customer stories Customer stories Customer stories Powered by Claude Powered by Claude Powered by Claude Service partners Service partners Service partners Availability Availability Availability Check fil

    匹配字段正文