Failed to crawl the webpage. (English version not available)

Jurassic Park 中呈现的技术，是了解九十年代早期计算世界的一个有趣窗口；片方为了追求真实感，专门使用了当时的高端硬件。控制室是这一展示的核心，摆放着来自 Silicon Graphics 和 Apple 的一系列机器。比如，Dennis Nedry 那杂乱的工位上有多台 Macintosh 和一台 SGI IRIS Crimson，而 Ray Arnold 更为整洁的桌面则放着一台 SGI R4000 Indigo 。制作组不惜借入价值数百万美元的设备，用当时最尖端的装备来布置布景，让场景看起来更可信。

控制室造型的一个标志性元素是 Thinking Machines CM-5 超级计算机，那些由数千颗闪烁红灯组成的面板极具视觉冲击力；在拍片时它们是世界上最强的计算系统之一。虽然面板上的灯光模式基本只是视觉效果，并不代表真实数据，但它们强化了影片的高科技氛围。还有诸如巨型 20 英寸 SuperMatch 20‑T 显示器和专用的 SGI Granite 键盘等专业级硬件，让影片更贴近 1993 年真实的工作站环境。

片中展示的软件界面既有实用道具，也有制作手法的巧妙运用。比如 Nedry 与联系人之间的视频通话并非实时，而是通过 QuickTime 播放的预录视频——能看见播放控制栏就是证据。 Lex Murphy 使用的著名 3D 文件浏览器其实是 SGI 为 IRIX 开发的实验性软件 fsn 。屏幕上出现的代码也是真实的，经鉴定为来自 Macintosh Programmers Workshop 的 Pascal 代码，说明即便是手册和源文件这样的细节，制作团队也精心挑选，以满足懂技术的观众。

影片还展示了当时前沿的便携设备。 Nedry 桌上的可折叠 PDA Motorola Envoy 就很特别——拍摄时它尚未商业化，能出现在片中是因为设备设计师与导演 Steven Spielberg 的一次偶遇，因而借来了稀有原型机。制作组对细节的讲究，从硬件规格到 Nedry 书架上的技术书，都体现出 Jurassic Park 在再现当时技术面貌上的用心良苦。 The technology featured in Jurassic Park is a fascinating look into early nineties computing, reflecting a production that prioritized authenticity by utilizing high-end hardware of the era. The control room serves as the epicenter of this display, housing an impressive collection of machines from Silicon Graphics and Apple. Dennis Nedry's chaotic workspace, for instance, features multiple Macintoshes and an SGI IRIS Crimson workstation, while Ray Arnold's more organized desk hosts an SGI R4000 Indigo. The production team went to great lengths to ensure these settings felt genuine, loaning millions of dollars worth of equipment to outfit the set with gear that was state-of-the-art at the time.

A central element of the control room's aesthetic is the presence of Thinking Machines CM-5 supercomputers. These machines, recognizable by their iconic panels of thousands of blinking red LEDs, were among the most powerful computing systems in the world when the movie was produced. While the light patterns on these machines were essentially random eye candy rather than meaningful data, their inclusion underscored the film's commitment to high-tech realism. Other professional-grade hardware, such as the massive 20-inch SuperMatch 20-T monitors and specialized SGI Granite keyboards, further grounded the film in the realities of professional workstation environments from 1993.

The software and interface designs shown on these machines often blended practical props with clever production tricks. For instance, the video conference call between Nedry and his contact was not a live feed but a pre-recorded sequence played through the QuickTime video player, a detail betrayed by the visible playback controls. Similarly, the famous 3D file explorer used by Lex Murphy was an actual experimental software tool called fsn, developed by SGI for the IRIX operating system. The code displayed on screen was also authentic, identified as Pascal from the Macintosh Programmers Workshop, proving that even minor details on props like manuals and source files were carefully curated to satisfy tech-savvy viewers.

Beyond the major workstations, the film showcased portable technology that was cutting-edge for the period. The Motorola Envoy, a foldable PDA found on Nedry's desk, is a notable inclusion because it was not yet commercially available when the movie was filmed. Its presence was the result of a chance encounter between the device's designer and director Steven Spielberg, leading to the use of a rare prototype. This commitment to detail, extending from the hardware specs down to the specific technical books on Nedry's shelf, highlights how deeply Jurassic Park was designed to represent the technological landscape of its time.

Tailscale 维护着一份公开记录，列出影响其客户端和服务基础设施的安全通知与漏洞。该记录透明呈现了以往事件，涵盖从轻微缺陷到涉及权限提升或潜在未经授权数据访问的严重问题。对每个漏洞，Tailscale 都说明了具体影响、受影响的平台或版本，以及用户应采取的修复步骤以保持网络安全。

这些通告反复强调保持 Tailscale 客户端为最新版本的重要性。例如，若干与 Tailscale Serve 、 Tailscale SSH 及本地 Web 界面权限相关的漏洞，均已通过升级到指定发布版本得到修复。更新通常修复路径处理错误、命令参数过滤不当或逻辑缺陷等问题，否则可能导致绕过访问控制或无意中以提升权限执行命令。

公告还凸显了在 macOS 、 Linux 、 Kubernetes 及云托管基础设施等多样环境中维护安全网络连接的复杂性。部分问题涉及时具体组件（如 Kubernetes operator 或 DERP server mesh），技术失误可能导致信息泄露或拒绝服务风险。 Tailscale 强调，许多漏洞通过软件更新修补，也有通过服务端改动或调整协调平面逻辑来缓解的情况。

除了技术缺陷，安全历史还涉及与身份提供商和共享邮箱域相关的运营风险。例如，公司解决了 tailnets 与电子邮件域映射的问题，新增了用户与设备审批等功能，以防止意外访问。公告提醒管理员审计配置，例如核查状态目录设置、审阅访问控制策略并监控审计日志以发现异常行为。

报告中对外部安全研究人员和社区成员发现并上报漏洞的贡献表示感谢。公司通过清晰记录事件并明确是否需要用户采取措施，体现了对透明度的承诺。维护该档案为 IT 与安全负责人评估平台历史并确保部署抵御已知威胁提供了重要参考。 Tailscale maintains a public record of security notifications and vulnerabilities affecting its client and service infrastructure. These bulletins provide transparency regarding past incidents, ranging from minor bugs to critical issues involving privilege escalation or potential unauthorized data access. For each reported vulnerability, the company outlines the specific impact, the affected platforms or versions, and the necessary remediation steps for users to maintain a secure networking environment.

A common theme across many of these disclosures is the importance of keeping the Tailscale client updated to the latest version. Several vulnerabilities, such as those related to Tailscale Serve, Tailscale SSH, and local web interface permissions, were resolved by upgrading to specific release versions. These updates often address issues like incorrect path handling, improper command argument sanitization, or logic flaws that could allow users to bypass established access control lists or execute commands with unintended elevated privileges.

The bulletins also highlight the complexities of managing secure network connectivity across diverse environments, including macOS, Linux, Kubernetes, and cloud-hosted infrastructures. Some issues involved specific components like the Kubernetes operator or the DERP server mesh, where technical oversights led to potential information disclosure or denial-of-service risks. Tailscale emphasizes that while many of these vulnerabilities were addressed through software updates, others were mitigated via server-side changes or by refining the logic within their coordination plane.

Beyond technical bugs, the security history covers operational risks associated with identity providers and shared email domains. For instance, the company addressed concerns regarding how tailnets are mapped to email domains, implementing features like user and device approval to prevent unintended access. These bulletins serve as a reminder for administrators to audit their configurations, such as ensuring proper state directory settings, reviewing access control policies, and monitoring audit logs for any unusual activity.

Throughout these reports, Tailscale credits external security researchers and community members for their role in identifying and reporting vulnerabilities. The company demonstrates a commitment to transparency by documenting these events clearly and offering specific guidance on whether user action is required. By maintaining this archive, Tailscale provides a vital resource for IT and security leaders to evaluate the platform's history and ensure their deployments remain resilient against known threats.

Vancouver Police Department (VPD) 是 Vancouver 市的主要执法机构，维护着一个综合性的数字平台，便于推动社区安全、公众报案与招聘工作。其使命核心是与社区携手，促进公共安全领域的卓越与创新，并遵循诚信、同情、问责、尊重与卓越的价值原则。该部门由 Chief Constable Steve Rai 领导，下辖按地理巡逻区域组织的宣誓警员与文职人员。

VPD 的重要职能之一是为犯罪预防和受害者提供可及的支持。部门为犯罪受害者提供危机干预与情感支持，并开展针对亲密伴侣暴力、人口贩运和心理健康等紧迫社会问题的专项举措。为保持透明并让公众知情，VPD 会发布犯罪统计数据，便于居民了解所在社区的趋势，从而采取主动措施保护财产与人身安全。

在需要警方协助时，VPD 明确区分紧急与非紧急情况：若面临直接危险或正在发生的犯罪，应拨打 911；对于轻微财产被盗、故意破坏或事后报案等情况，可使用非紧急报案系统。 VPD 还在总部及其他指定服务点通过公共服务柜台为公众办理 Police Information Checks 、指纹采集等行政事务。

除了日常执法外，VPD 也积极参与大型社区活动，例如担任 FIFA World Cup 2026 的主责警务机构。部门高度重视青少年外展项目，旨在为青少年提供远离犯罪、帮派与毒品的替代路径。为改善与公众的沟通，VPD 推出了 VPD Connect 移动应用，用于发布实时安全警报与相关信息。

招募工作对部门至关重要，VPD 定期为宣誓警官与文职岗位寻觅合格且多元化的候选人，并通过线上线下信息会以及像 "Run with a Recruiter" 这样的参与活动，帮助候选人准备警务训练的体能要求。通过在社区参与、运作透明与职业发展等方面的综合努力，Vancouver Police Department 致力于适应这座多元文化国际港口城市不断变化的需求。 The Vancouver Police Department (VPD) serves as the primary law enforcement agency for the City of Vancouver, maintaining a comprehensive digital platform to facilitate community safety, public reporting, and recruitment. At the core of its mission is the goal of partnering with the community to foster excellence and innovation in public safety, operating under the guiding principles of integrity, compassion, accountability, respect, and excellence. The department is led by Chief Constable Steve Rai, who oversees a workforce comprised of both sworn officers and civilian staff organized into geographical patrol areas.

A significant component of the VPD's operation involves providing accessible resources for crime prevention and victim support. The department offers specialized services for victims of crime, including crisis intervention and emotional support, and maintains specific initiatives to address pressing social issues such as intimate partner violence, human trafficking, and mental health challenges. To maintain transparency and inform the public, the VPD publishes crime statistics, enabling residents to track trends in their neighborhoods and providing them with the necessary information to take proactive measures in protecting their property and personal safety.

For those requiring police assistance, the department clearly distinguishes between emergency and non-emergency situations. In cases of immediate danger or active crimes, the public is directed to call 911. Conversely, the department maintains a non-emergency reporting system for incidents such as minor property theft, vandalism, or retrospective reports of crime. The VPD also handles administrative functions for the public, including police information checks and fingerprinting services, through its public service counters located at headquarters and other designated facilities.

Beyond routine enforcement, the VPD is actively involved in large-scale community events, such as serving as the lead police agency for the FIFA World Cup 2026. The department also places a strong emphasis on youth outreach programs, which are designed to provide alternatives to criminal involvement, gangs, and drug-related activities. To improve communication with the public, the department has introduced the VPD Connect mobile app, which serves as a tool for delivering real-time safety alerts and information.

Recruitment remains a vital activity for the department, which regularly seeks qualified, diverse candidates for both sworn officer positions and civilian roles. The VPD supports this process by hosting in-person and virtual information sessions, as well as unique engagement opportunities like "Run with a Recruiter," which helps candidates prepare for the physical requirements of police training. Through these combined efforts in community engagement, operational transparency, and professional development, the Vancouver Police Department strives to adapt to the evolving needs of its multicultural, international port city.

Microsoft 发布了大规模软件更新，修复了其 Windows 生态系统及其他产品中创纪录的 570 个安全漏洞。该数字几乎是上个月的三倍，微软将增长归因于在漏洞发现流程中引入了人工智能。其中近 60 个漏洞被评为关键级别，攻击者在极少用户交互的情况下即可远程控制设备。

此次更新还修补了三个零日漏洞，其中两个已在野外被利用。值得关注的问题包括 Active Directory Federation Services 中的权限提升漏洞和 Microsoft SharePoint 中的另一个漏洞。此外，Windows BitLocker 中的一个漏洞可能允许拥有物理访问权限的攻击者绕过安全保护并访问加密数据。

专家指出，人工智能在加速漏洞发现与修复的同时，也让攻击者更快地开发利用手段，这使得以人为中心的指标（例如 Microsoft 的 exploitability index）可能不再能充分反映实际威胁程度。研究显示，AI 模型能够为先前被认为不太可能被武器化的漏洞生成概念验证型利用代码。

整个软件行业也在经历类似变化，厂商难以适应越来越频繁的安全公告节奏。例如，Adobe 已改为每月两次发布更新，其他大型科技公司也在更频繁地推送补丁。这一转变凸显了软件维护方式的系统性变化，公司正努力适应 AI 驱动漏洞检测带来的快速节奏。

鉴于本次补丁数量庞大，安全专家建议用户谨慎操作。尽管更新对长期安全至关重要，但如此大规模的变更偶尔会引发系统稳定性问题。建议用户先备份数据，并可考虑等待几天再在关键系统上应用更新，以观察是否出现意外问题。 Microsoft has released a massive set of software updates, patching a record-breaking 570 security flaws across its Windows ecosystem and other products. This figure is nearly triple the count from the previous month, a jump that the company attributes to the integration of artificial intelligence in its vulnerability discovery processes. Nearly 60 of these newly identified bugs are classified as critical, allowing potential attackers to gain remote control over devices with minimal user interaction.

The update package also addresses three zero-day vulnerabilities, two of which are currently being exploited in the wild. Among the notable issues is an elevation of privilege flaw in Active Directory Federation Services and another in Microsoft SharePoint. Furthermore, a vulnerability in Windows BitLocker has been identified that could grant an attacker with physical access to a device the ability to bypass security features and access encrypted data.

Experts point out that while AI is accelerating the discovery and remediation of these flaws, it is also providing attackers with the same capabilities to quickly develop exploits. This creates a challenging environment where the traditional human-centered metrics, such as Microsoft's own exploitability index, may no longer adequately reflect the actual threat level. Research has shown that AI models can successfully generate proof-of-concept exploits for vulnerabilities that were previously deemed unlikely to be weaponized.

The broader software industry is experiencing a similar trend as manufacturers struggle to keep up with an increasing cadence of security bulletins. Adobe, for instance, has moved to a twice-monthly update schedule, while other major tech firms are also pushing out patches more frequently. This shift highlights a systemic change in how software maintenance is conducted, as companies adapt to the rapid pace of AI-driven vulnerability detection.

Given the sheer volume of patches included in this release, security professionals suggest that users exercise caution. While updating is essential for long-term security, the complexity of such a large batch of changes can occasionally lead to system stability issues. It is recommended that users back up their data and consider waiting a few days to ensure that the updates do not cause unforeseen technical problems before applying them to critical systems.

GitHub 为 Dependabot 的版本更新引入了默认冷却期，以加强供应链安全。今后，Dependabot 在某个软件包在其注册表发布新版本后，会至少等待三天才为该版本创建 Pull Request 。此更改将在 github.com 上对所有受支持的生态系统默认生效，并计划纳入 GitHub Enterprise Server 3.23 。

此举主要是为了降低供应链攻击风险。新版本有时可能被篡改或不稳定，攻击者常利用发布后短暂的时间窗口注入恶意代码。通过设置短暂的等待期，GitHub 给社区反馈与安全信号留下时间，从而减少开发者在版本刚发布时就不经意合并潜在的恶意或有问题代码的可能性。

需要说明的是，该默认仅适用于普通版本更新，安全更新不受影响，仍会立即开启以确保关键漏洞及时修补。 GitHub 强调这是主动的安全措施，而非限制性做法，用户对工作流仍保有完全控制权。

若团队有不同偏好，可以在 .github/dependabot.yml 配置文件中修改 cooldown 选项来自定义或完全禁用冷却期，便于在提升安全性与满足项目需求或部署节奏之间取得平衡。 GitHub has introduced a new default cooldown for Dependabot version updates to enhance supply chain security. Moving forward, Dependabot will wait at least three days after a new package release becomes available on its registry before creating a pull request for that version update. This change is being implemented by default across all supported ecosystems on github.com and is scheduled to be included in GitHub Enterprise Server version 3.23.

The primary motivation behind this update is to mitigate risks associated with supply chain attacks. New software releases can sometimes be compromised or unstable, and malicious actors often exploit the brief period after a release to inject harmful code. By enforcing a short waiting period, GitHub provides enough time for community feedback and security signals to emerge, which helps ensure that developers do not inadvertently merge potentially malicious or broken code as soon as it is published.

It is important to note that this new default applies exclusively to standard version updates. Security updates remain unaffected and will continue to be opened immediately to ensure that critical vulnerabilities are patched without delay. GitHub emphasizes that this is a proactive security measure rather than a restrictive one, as users retain full control over their workflows.

For teams that prefer a different approach, the cooldown period can be customized or entirely disabled by modifying the cooldown option within the .github/dependabot.yml configuration file. This allows developers to balance the need for enhanced security with their specific project requirements or deployment cycles.

在 Go 中使用 HTMX，可以在尽量减少自定义 JavaScript 的同时，为 Web 应用增加交互性，并保持服务端 HTML 渲染的可靠性。借助 Go 的 html/template 包，开发者能构建流畅、近似原生应用的体验。常见做法是用一个共享的基础布局加若干小的局部模板（partials）来组织模板，这样应用就能根据用户交互的上下文，返回完整页面或针对性的 HTML 片段来响应 HTMX 请求。

实现时通常会用 embed.FS 将 HTML 和静态资源嵌入到 Go 二进制里，简化部署。一个自定义的渲染器类型可以在启动时解析模板，运行时由处理器动态克隆并扩展模板集合。这种模块化让代码更干净、遵循 DRY 原则，局部模板可以在应用不同位置复用，比如搜索结果中的表格行或小型 UI 组件（如图片）。

处理 HTMX 请求的逻辑需要注意 HTTP 头。因为 HTMX 请求会包含 HX-Request 头，Go 的处理器可以据此判断是返回完整页面还是局部片段；同时设置 Vary: HX-Request 可以让中间缓存正确区分这两类响应。为避免历史记录恢复出错，应配置 HTMX 避免不当的历史缓存，这样用户用后退按钮返回时能看到预期的完整页面，而不是损坏的局部内容。

还要细化重定向和错误处理。标准的 3xx 重定向可能会在 HTMX 处理前被浏览器拦截，因此用带 2xx 状态码的 HX-Redirect 头能更好地由服务端控制导航。并且通过自定义 HTMX 的响应处理策略，可以定义不同状态码（例如用于验证错误的 422 或用于服务器问题的 500）应如何影响 DOM，确保错误信息清晰呈现而不是静默失败。

对于更复杂的需求——例如按页面定制布局或需要获取浏览器当前 URL——上述模式依然高度可扩展。禁用不必要的功能（如客户端历史缓存和属性继承）可以让应用更可预测、更安全。把配置和模板管理的复杂性集中到 Go 擅长的服务端，会得到一个健壮且易维护的架构。 Using HTMX with Go is an effective way to add interactivity to web applications while minimizing custom JavaScript and maintaining the reliability of server-side HTML rendering. By leveraging Go's html/template package, developers can create smooth, app-like experiences. The core approach involves structuring templates with a shared base layout and smaller partials, allowing the application to respond to HTMX requests with either a full page or a specific, targeted HTML fragment depending on the context of the user interaction.

To implement this, it is standard practice to embed HTML and static assets directly into the Go binary using embed.FS, which simplifies deployment. A custom renderer type can manage template parsing at startup, allowing handlers to dynamically clone and extend template sets. This modularity enables developers to write clean, DRY code where partials can be reused across different parts of the application, such as table rows in a search interface or small UI components like images.

Handling the logic for HTMX requests requires careful attention to HTTP headers. Because HTMX requests include an HX-Request header, Go handlers can detect them to determine whether to return a full page or just a partial snippet. Similarly, setting the Vary: HX-Request header ensures that intermediate caches correctly distinguish between these different response types. This setup also addresses navigation, where configuring HTMX to avoid improper history restoration ensures that users returning to a page via the back button receive the expected full-page experience rather than a broken partial.

Refining the user experience involves managing redirects and errors. Since standard 3xx redirects can be intercepted by the browser before HTMX sees them, using an HX-Redirect header with a 2xx status code allows for better control over navigation from the server. Furthermore, customizing HTMX response handling settings allows developers to define how different status codes—such as 422 for validation errors or 500 for server issues—should affect the DOM, ensuring that error messages are clearly presented to the user rather than failing silently.

For more complex requirements, such as handling page-specific layouts or needing to know the browser's current URL, the patterns described remain highly extensible. By disabling unnecessary features like client-side history caching and attribute inheritance, developers can maintain a more predictable and secure application environment. This disciplined approach to configuration and template management results in a robust, maintainable architecture that keeps the complexity centered on the server, where Go excels.

Mindgard 的安全研究人员在 Cursor 的 AI 辅助开发环境中发现了一个关键漏洞，该漏洞可导致任意代码执行。漏洞非常简单：当开发者在 Cursor 中打开项目时，应用会自动扫描是否存在 Git 可执行文件。如果仓库根目录下存在名为 git.exe 的恶意文件，Cursor 会立即执行该文件，且不会弹出任何提示、警告或征得用户许可。只要仓库保持打开，该执行过程会反复触发，给用户带来持续的安全风险。

尽管 Mindgard 于 2025 年 12 月 15 日报告了该漏洞并通过多种渠道跟进，但在超过七个月的时间里问题一直未得到修补。 Cursor 在一段时间没有回应后才承认问题，但公司既未提供更新、也未展示修复进展或采取保护用户的措施，反而在此期间发布了数十次软件更新。鉴于该漏洞性质直观且厂商缺乏应对，研究人员最终选择全面公开披露。

该漏洞影响严重：Cursor 被数百万开发者和数千家企业广泛使用，而利用该问题几乎只需打开一个被植入恶意文件的项目，凸显出现代快速演进的开发平台中存在的固有安全隐患。研究人员强调，对此类工具的信任应通过透明的行为和对安全缺陷的主动修复来建立，而不仅仅凭借工具的实用性。

针对缓解措施，受 Windows 管理的系统可临时使用 AppLocker 或 Windows App Control 等工具，限制工作区目录中指定可执行文件的运行。对个人用户，研究人员建议仅在隔离环境中打开不受信任的仓库，例如在虚拟机或 Windows Sandbox 中操作。对这些权宜之计的依赖凸显了当厂商未尽职责时，用户需要被告知风险并自行采取防护措施的必要性。

最终，此次披露反映出安全生态中一个更广泛且令人担忧的变化：传统的漏洞披露流程难以跟上以 AI 为重点、节奏快且高增长的公司环境。 Mindgard 发布这些细节，旨在帮助开发者和组织就其安全态势做出明智决策。研究人员认为，当一家公司停止沟通并将用户暴露在已知威胁中时，公开披露成为防止进一步风险的最后手段。 Security researchers at Mindgard have identified a critical vulnerability in the Cursor AI-assisted development environment that allows for arbitrary code execution. The flaw is remarkably simple. When a developer opens a project in Cursor, the application automatically scans for Git binaries. If a malicious file named git.exe is present in the root of the repository, Cursor executes it immediately without any user prompts, warnings, or approval. This process occurs repeatedly while the repository is open, posing a persistent security risk to users.

Despite being reported by Mindgard on December 15, 2025, and followed up on through various channels, the vulnerability remained unpatched for over seven months. While Cursor eventually acknowledged the issue after a period of unresponsive communication, the company failed to provide updates, demonstrate progress on a fix, or protect its users, even as it continued to release dozens of software updates. This lack of engagement, despite the straightforward nature of the bug, prompted the researchers to move to full public disclosure.

The implications of this vulnerability are significant, as Cursor is a widely adopted tool used by millions of developers and thousands of companies. Because the exploitation of this bug requires nothing more than opening a tainted project, it serves as a stark reminder of the security risks inherent in modern, rapidly evolving development platforms. The researchers emphasize that trust in such tools must be earned through transparent behavior and a proactive commitment to addressing security flaws, rather than just utility.

For users on Windows-managed systems, temporary mitigation involves using tools like AppLocker or Windows App Control to restrict the execution of specific binaries within workspace directories. On consumer systems, researchers advise opening untrusted repositories only within isolated environments such as virtual machines or Windows Sandbox. The reliance on such workarounds highlights the necessity for users to be informed of risks when vendors fail to fulfill their responsibilities.

Ultimately, this disclosure highlights a broader, troubling shift in the security landscape where traditional disclosure pipelines struggle to keep up with the fast-paced, high-growth environment of AI-focused companies. By releasing these details, Mindgard aims to ensure that developers and organizations can make informed decisions about their security posture. The researchers contend that when a company stops communicating and leaves users exposed to known threats, public disclosure becomes the only remaining option to prevent further risk.

PrismML 宣布发布 Bonsai 27B，这是一款基于 Qwen3.6 的强大多模态模型，标志着端侧 AI 的一个重要里程碑。此前已有模型表明低位权重可以有效，但 Bonsai 27B 是同类能力中首个能在手机上运行的版本。这一成就将多步推理、结构化工具调用和复杂的代理循环等高级功能带到本地硬件上——这些功能以前因 27B 参数模型所需的大量内存而难以在设备上实现。

本次发布包含两个针对不同性能与资源需求的版本。 Ternary Bonsai 27B 采用三值权重，使得每个权重的有效位数约为 1.71 比特，占用约 5.9 GB 内存，旨在为普通笔记本提供高质量的推理与代理能力。 1-bit Bonsai 27B 采用二值权重，每个权重有效位数约为 1.125 比特，将模型体积缩至约 3.9 GB，足够在 iPhone 17 Pro 的内存限制内运行，体现了在单位内存中提升智能密度的重大突破。两款模型均支持 262K-token 的上下文窗口，并通过推测性解码提高运行速度。

在性能上，Bonsai 27B 系列与全精度模型非常接近。在覆盖 15 项基准的测试中，Ternary 模型保持了基线能力的 95%，1-bit 版本保持了约 90% 。尤其是数学、编程和工具调用等核心能力仍然很强。这种能力保留使得模型可以处理需要持续且精准推理的复杂代理工作流，有效缩小云端任务与本地处理之间的差距。

向本地执行的转变解决了云端 AI 的若干限制，尤其对在单次工作流中执行数百步操作的代理来说意义重大。设备原生运行可以消除网络请求带来的延迟与费用，同时确保敏感用户数据（如私人文件和屏幕内容）保留在本地硬件。该架构还支持混合部署：私密或高频任务在本地处理，而更大算力需求的前沿任务则可上云执行。

Bonsai 系列的方法论把"智能密度"作为未来 AI 发展的关键指标，通过在有限内存中尽量装入更多能力，PrismML 致力于让高端 AI 在更广泛的场景中可用。这些模型以 Apache 2.0 License 提供，并针对通过 MLX 运行的 Apple silicon 以及使用定制低位 CUDA 内核的 NVIDIA GPU 进行了优化。随着公司持续改进压缩技术，预计会不断推动日常设备上 AI 能力的边界。 PrismML has announced the release of Bonsai 27B, a powerful new multimodal model based on Qwen3.6 that marks a significant milestone in on-device AI. While previous models have demonstrated that low-bit weights could be effective, Bonsai 27B is the first of its capability class to operate on a mobile phone. This achievement brings advanced features such as multi-step reasoning, structured tool calls, and sophisticated agentic loops to local hardware, which were previously impractical due to the massive memory requirements typically associated with 27B-parameter models.

The release includes two distinct variants designed for specific performance and footprint needs. The Ternary Bonsai 27B, which uses three-value weights to achieve an effective 1.71 bits per weight, occupies 5.9 GB. It is built to deliver high-quality reasoning and agentic capabilities on standard laptops. The 1-bit Bonsai 27B, using binary weights for an effective 1.125 bits per weight, shrinks the footprint to just 3.9 GB. This specific version is small enough to fit within the memory constraints of an iPhone 17 Pro, representing a notable breakthrough in intelligence density. Both models support a 262K-token context window and leverage speculative decoding for enhanced speed.

In terms of performance, the Bonsai 27B variants remain remarkably close to their full-precision counterparts. Across a broad 15-benchmark suite, the Ternary model retains 95% of the baseline intelligence, while the 1-bit version keeps 90%. Notably, core capabilities like mathematics, coding, and tool calling remain highly effective. This level of retention allows the models to handle complex agentic workflows where sustained, accurate reasoning is essential, effectively bridging the gap between cloud-dependent tasks and local processing.

The shift toward local execution addresses key limitations of cloud-based AI, particularly for agents that perform hundreds of steps in a single workflow. By running natively on the device, the system removes the latency and costs associated with network requests while ensuring that sensitive user data, such as private files and screen content, remains on the hardware. This architecture also supports hybrid deployments, where private or high-frequency tasks are handled locally while more demanding, frontier-level tasks are sent to the cloud.

The underlying methodology of the Bonsai series emphasizes intelligence density as a critical metric for future AI development. By focusing on how much capability can be packed into a limited memory footprint, PrismML is working to make high-end AI accessible across a wide range of environments. The models are available under the Apache 2.0 License and are optimized for both Apple silicon via MLX and NVIDIA GPUs using custom low-bit CUDA kernels. As the company continues to refine its compression techniques, it expects to keep pushing the boundaries of what is possible on everyday devices.

Tower of Babel 的故事常被解读为对傲慢和人类野心危险的警示，但同样可以用来说明集体成功依赖于团结与协作。在圣经的记载中，建造者拥有共同的语言，使他们能够同步行动，完成任何单个个体都无法完成的壮举。当他们彼此理解的能力被剥夺时，塔的建设就停了下来。这凸显出，技术与文明的进步从根本上依赖于一群人维持统一目标和共同话语的能力。

在现代软件开发的语境下，人工智能辅助编程常被宣称能提升个人生产力并加快代码产出速度。毋庸置疑，AI 代理确实让开发者更有能力修改代码库，但大规模软件项目的主要约束很少是代码产出速度。相反，这些项目的成败更多取决于团队能否有效地维持对系统的共享理解。这种内部语言包括约定的概念、架构边界以及系统为何这样设计的理由。

传统上，这种共享理解是通过一定程度的摩擦来维持的。当开发者需要修改系统中复杂的部分时，他们必须与他人沟通、提出问题，并将自己的意图与那些拥有或依赖该代码的人同步。尽管这个过程有时效率不高，但这种摩擦起着关键作用：它迫使开发者相互学习，确保所有相关人员对系统的运作保持一致。这种协作过程充当了同步机制，把集体知识嵌入到开发团队的人际网络中。

如今，AI 代理威胁着消除这种必要的摩擦。开发者可以利用代理在孤立的情况下实施变更，常常无需咨询同事或深入理解更广阔的架构。每一次独立修改看起来都合理、能通过必要的测试，但累积起来的结果却是在侵蚀共享的架构话语。由于代理可以按需生成解释，开发者得以在各自的孤岛式、局部化方式中工作，实际上绕开了维持对整个项目连贯统一心智模型的需要。

这就产生了一种与圣经中关于 Tower of Babel 的神话不同的新型、令人迷惑的局面。在原故事中，失去共同语言导致了立即的失败和建设的停止。而在现代的人工智能辅助工程中，建设并没有停止。因为代码仍能被不知疲倦的 AI "翻译者"管理，即便底层的人类对架构的理解已实际崩溃，项目仍会继续扩展。塔在不断升高，但曾经维系其完整性的凝聚力已被孤立的、由机器驱动的局部改动取代，这些改动掩盖了参与者已无法真正统筹系统全局的事实。 The story of the Tower of Babel is often interpreted as a warning about pride and the dangers of human ambition, but it serves just as effectively as an illustration of how collective success relies on unity and coordination. In the biblical account, the builders possessed a shared language that allowed them to synchronize their efforts and achieve feats impossible for any single individual. When their ability to understand one another was stripped away, the construction of the tower halted. This highlights that technological and civilizational progress is fundamentally tied to the ability of a group to maintain a unified purpose and a common vocabulary.

In the modern context of software development, AI-assisted programming is often touted as a way to boost individual productivity and accelerate the pace at which code is produced. While it is undeniable that AI agents make developers more capable of modifying a codebase, the primary constraint on large-scale software projects has rarely been the speed of code production. Instead, these projects thrive or struggle based on how effectively teams can maintain a shared understanding of their system. This internal language consists of agreed-upon concepts, architectural boundaries, and the reasoning behind why a system is shaped the way it is.

Traditionally, this shared understanding was maintained through a degree of friction. When a developer needed to modify a complex part of a system, they were required to engage with other people, ask questions, and synchronize their intentions with those who owned or depended on that code. While this process was sometimes inefficient, that friction served a critical purpose. It forced developers to learn from one another and ensured that everyone involved remained aligned on how the system worked. This collaborative process acted as a synchronization mechanism, embedding collective knowledge into the human network of the development team.

Today, AI agents threaten to remove that essential friction. Developers can now utilize agents to implement changes in isolation, often without needing to consult their peers or gain a deep, contextual understanding of the broader architecture. Each individual modification might appear reasonable and pass all necessary tests, but the collective result is an erosion of the shared architectural language. Because agents can generate explanations on demand, developers can work in siloed, localized ways, effectively bypassing the need to maintain a coherent, unified mental model of the entire project.

This creates a new, disorienting scenario that diverges from the biblical myth of Babel. In the original story, the loss of a common language led to immediate failure and the cessation of construction. In modern AI-assisted engineering, however, the construction does not stop. Because the code can still be managed by tireless AI translators, the project continues to grow even after the underlying human understanding of the architecture has effectively collapsed. The tower keeps rising, but the cohesion that once held it together has been replaced by isolated, machine-driven local changes that mask the fact that the humans involved can no longer truly reason about the system as a whole.