Google 发布 Gemini 4 Argon，HN 条目拿到 1328 分，是当日最高分^[1]；同日 X 上也有开发者称它「对世界有惊人的自然理解，能从头原理出发做很多事」并举了生成示例^[6]。它要解决的是模型能力评估的日常化：与其看基准分数，用户更愿意用随手能验证的任务（画一只骑自行车的鹈鹕）来感知模型的物理常识与组合能力。做法上发布方提供模型页与 API，社区自发形成测试集。对做模型选型的团队，参考价值是把「社区惯例测试」纳入早期试用，比等正式评测更快建立直觉；限制是这类主观示例不能替代基准，且 Google 官方页面的可用范围、定价与地区限制（此前一篇报道提到欧盟部分功能受限）都需要另行核对。

Google released Gemini 4 Argon, whose HN thread reached 1328 points as the day's top item^[1]; on X a developer called it having “amazing natural understanding of the world” able to do many things from first principles, with a generated example^[6]. It addresses how model capability is assessed day to day: rather than benchmark scores, users prefer tasks they can verify instantly to sense physical common sense and composition. The vendor ships a model page and API while the community builds its own tests. For model-selection teams the reference is to fold community-style spot tests into early trials before formal evaluations land; the limit is that such subjective examples do not replace benchmarks, and availability, pricing and regional restrictions must be checked separately.

瑞典 TV4 的节目《The Challenge》因技术故障导致已剪好的素材被抹掉而推迟到 2027 年播出，HN 上 148 分，报道把原因指向 AI Agent 的误操作^[21]。它要解决的是生成式工具介入专业后期流程后的破坏性风险：剪辑素材往往没有可用的版本管理，而 Agent 的批量操作是「全有或全无」——一次误删就可能让数周工作归零。做法上媒体以推迟整季来止损，说明缺少可回滚的素材管理。对做内容生产与 Agent 自动化的团队，参考价值是在把写操作交给 Agent 前，先确保素材有版本控制与快照，否则能力再强也不敢用；限制是报道来自瑞典媒体转述、技术细节（是模型误判还是流程脚本缺陷）尚未确认，不宜把责任直接归给模型。

Swedish broadcaster TV4 postponed its show “The Challenge” to 2027 after a technical failure erased edited footage, scoring 148 points on HN, with reports pointing to an AI agent's erroneous operation^[21]. It addresses the destructive risk of generative tooling inside professional post-production: edit material often has no usable version control, and an agent's bulk operation is all-or-nothing — one mistake can erase weeks of work. The broadcaster absorbed the loss by pushing the season. For content production and agent automation teams the reference is to guarantee versioning and snapshots before handing write operations to an agent, or capability goes unused; the limit is that this comes from a second-hand report and the technical cause is unconfirmed.

这条新闻有两个信息源层次：地方媒体与彭博社报道一名 OpenAI 工程师因在行政会议室暗藏录音设备被捕^[3]，而起诉文件显示其使用的是一块 Seeed Xiao ESP32-S3 Sense 开发板^[4]；OpenAI 官方 X 账号也发帖说明在办公室发现了隐蔽的录音装置，并用自家模型分析该设备^[5]（该帖 7.0 万+ 赞）。它要解决的是内部威胁与廉价硬件的交汇问题：几十美元的开发板加上云端模型，就能构成过去需要专业设备的能力，物理安全与技术审计必须一起做。对做企业安全与办公场所管理的团队，参考价值是把「低成本可联网录音设备」纳入物理安全检查清单，并限制内部设备的网络出站；限制是案件仍在司法程序中，动机与责任尚未认定，报道细节也未必完整。

This story has three layers: local and Bloomberg reporting that an OpenAI engineer was arrested over a hidden recording device in an executive conference room^[3], filings showing the device was a Seeed Xiao ESP32-S3 Sense dev board^[4], and OpenAI's official X account stating that it discovered a sophisticated hidden recording device in its offices and used its own models to analyse it^[5] (70k+ likes). It addresses the intersection of insider threat and cheap hardware: a board costing tens of dollars plus cloud models now provides capability that once required professional equipment, so physical security and technical audit must move together. For corporate security teams the reference is to add low-cost networked recorders to physical security checklists and restrict internal device egress; the limit is that the case is still in process.

彭博法律报道（经 Techmeme 整理），在苹果诉 OpenAI 商业秘密案中，美国法官部分批准了苹果的加速取证动议，命令 OpenAI 交出设备取证镜像等材料^[8]。它要解决的是模型公司内部信息与外部合作方之间的边界：当人员流动与产品路线交叉时，商业秘密争议会要求交出设备级证据，而这类取证对企业内部治理与法务流程提出真实要求。做法上法院以证据开示程序推进事实认定。对做技术合作与人员流动管理的团队，参考价值是设备与数据的留存策略会影响未来诉讼的成本与结果，应提前与法务对齐；限制是案件尚未判决，报道未披露争议的技术细节与涉嫌范围，结论需要等审理进展。

Bloomberg Law, summarised by Techmeme, reports that in Apple's trade-secrets suit against OpenAI, a US judge partly granted Apple's motion for expedited discovery, ordering OpenAI to hand over forensic images of devices and other documents^[8]. It addresses the boundary between a model company's internal information and its external partners: when people and product roadmaps overlap, trade-secret disputes demand device-level evidence, which places real requirements on internal governance and legal process. The court advances fact-finding through discovery. For teams managing technical partnerships and personnel movement the reference is that device and data retention policy shapes the cost and outcome of future litigation, so align with legal early; the limit is that the case is undecided.

Fortune 报道（经 Techmeme 整理），位于特拉华州的 Causa Prima 完成 2,000 万美元融资，其产品是面向 B2B 支付管理、账单争议解决等场景的「Agent 到 Agent」平台^[19]。它要解决的是企业间事务由 Agent 代理后的结算与仲裁问题：当双方都用 Agent 谈判、下单与对账，争议不再是人和人之间的邮件往来，而需要机器可读的证据与规则。做法上把支付与争议流程做成 Agent 可直接参与的结构化协议。对做 Agent 商业化的团队，参考价值是「Agent 之间怎么结算与举证」是接下来最实际的一层基础设施；限制是报道未披露客户与产品细节，B2B 场景的合规与责任划分复杂，落地周期可能比消费场景更长。

Fortune, summarised by Techmeme, reports that Delaware-based Causa Prima raised $20M for an agent-to-agent platform covering B2B payment management, billing dispute resolution and more^[19]. It addresses settlement and arbitration once enterprises delegate transactions to agents: when both sides use agents to negotiate, order and reconcile, disputes stop being human email threads and require machine-readable evidence and rules. The design turns payment and dispute flows into structured protocols agents can join. For agent commercialisation teams the reference is that how agents settle and evidence claims is the next practical infrastructure layer; the limit is that customers and product detail are undisclosed.

Techmeme 收录的细节补上了昨天那条 Anthropic 工程博客的量级：Scion 是 Anthropic 为单体仓库自研的 git 替代品，由 7 名工程师借助 Claude 在五个月内完成，同期每日合并量从 3 月的 2,500 次显著上升^[7]。它要解决的是Agent 提交量级下的版本控制瓶颈：当代码提交频率上升一个量级，既有工具在合并吞吐、评审与历史查询上都会成为瓶颈。做法上重写核心组件而不是继续在旧工具上打补丁。对做工程基础设施的团队，参考价值是当 Agent 把提交量推高一档时，值得认真评估「是改造还是重写」；限制是这条与 10-13 那期共用同一来源（今天新增的是人员、时长与合并量为量级信息），且单体仓库规模特殊，多数团队未必适用。

Details carried by Techmeme add scale to yesterday's Anthropic engineering post: Scion is Anthropic's in-house git replacement for its monorepo, built by seven engineers with Claude in five months, as daily merges rose substantially from 2,500 in March^[7]. It addresses version control under agent-scale commits: once commit frequency rises an order of magnitude, merging throughput, review and history queries all become bottlenecks. The approach rewrites core components rather than patching existing tools. For engineering infrastructure teams the reference is to seriously evaluate rebuild versus retrofit once agents push commit volume up a tier; the limit is that this shares its source with the 10-13 edition, adding only team size, duration and merge volume.

Google 开发者博客发布 Gemini 4 Argon for Robotics，HN 上获得 220 分^[2]。它要解决的是通用模型接机器人的最后一公里：机器人开发者需要的是低延迟、可处理视觉与空间关系、并能稳定输出结构化动作指令的接口，而不是通用对话能力。做法上在同一代模型下提供面向机器人场景的专门配置与文档。对做具身系统的团队，参考价值是大厂开始把机器人当作独立开发者群体来服务，这会显著降低感知与指令层的自研成本；限制是条目为开发者公告，未给出延迟、成功率与硬件兼容清单，真实机器人上的表现需要按任务实测。

Google's developer blog published Gemini 4 Argon for Robotics, scoring 220 points on HN^[2]. It addresses the last mile of connecting general models to robots: robot developers need low latency, vision and spatial reasoning, and reliably structured action commands rather than general conversation. The design ships a robot-oriented configuration and documentation within the same model generation. For embodied systems teams the reference is that large vendors now treat robotics as its own developer audience, which cuts the cost of building perception and instruction layers; the limit is that this is a developer announcement without latency, success-rate or hardware-compatibility details.

彭博社报道（经 Techmeme 整理），中国制造商在 2026 上半年占据了全球人形机器人出货量的 97%，背后是政府支持与庞大的供应链^[9]。它要解决的是机器人产业的制造集中度问题：人形机器人的瓶颈正在从算法转向「灵巧手、执行器与量产能力」，而供应链与成本优势会直接决定谁能先形成规模交付。做法上报道以出货数据与产业链布局呈现集中度。对做具身产品的团队，参考价值是选型与供应链策略必须正视制造集中度，硬件成本曲线由这一侧的产能决定；限制是统计口径（是否含教育与展示机型、如何界定「人形」）未公开，出货量也不等于稳定商用部署。

Bloomberg, summarised by Techmeme, reports that Chinese manufacturers captured 97% of global humanoid robot shipments in H1 2026, supported by government backing and a vast supply chain^[9]. It addresses manufacturing concentration in robotics: the bottleneck is shifting from algorithms to dexterous hands, actuators and production capacity, and supply-chain and cost advantages decide who reaches volume delivery first. The reporting presents concentration through shipment data and industrial layout. For embodied product teams the reference is that sourcing strategy must acknowledge this concentration, since the hardware cost curve depends on capacity on that side; the limit is that counting methodology is undisclosed and shipments are not the same as durable commercial deployment.

WSJ 报道（经 Techmeme 整理）称，自 2025 年 9 月以来，美国国防企业与创业公司已与台湾硬件公司签署至少 13 笔协议，目的是建立「不依赖中国」的无人机技术供应链^[10]。它要解决的是无人机与机器人供应链的地缘风险：电机、电池与飞控等关键部件的来源决定了产能与合规，而台湾的电子制造生态被视为替代路径。做法上以订单与合资把供应链转移到可信辖区。对做硬件采购与地缘风险规划的团队，参考价值是把「零部件的原产地」纳入设计阶段约束，而不是采购阶段才发现问题；限制是报道未披露交易金额与产能爬坡计划，实际替代能力需要几年时间验证。

The WSJ, summarised by Techmeme, reports that since September 2025 US defence companies and startups have signed at least 13 deals with Taiwanese hardware companies to build a drone technology supply chain free of Chinese dependence^[10]. It addresses geopolitical risk in drone and robotics supply chains: the origin of motors, batteries and flight controllers determines capacity and compliance, and Taiwan's electronics ecosystem is treated as the alternative. The approach shifts supply chains into trusted jurisdictions through orders and joint ventures. For hardware sourcing and geopolitical planning teams the reference is to treat component origin as a design-stage constraint rather than a procurement-time surprise; the limit is that deal values and capacity ramp are undisclosed.

Show HN 项目 Parsimony 是一个开源基准，用来衡量 AI 模型在处理任务时「额外增加了多少代码」^[11]（HN 69 分）。它要解决的是代码增量与收益脱节的问题：Agent 常常用大量新增代码解决问题，而新增代码本身带来维护、评审与缺陷半径的成本，此前缺少可比口径。做法上把「改动量」作为独立维度来测量，而不是只看是否通过测试。值得借鉴的是把「净增量」当作与正确性并列的指标纳入评测；限制是条目为早期项目，未给出覆盖的语言与任务范围，也没有与真实维护成本的关联验证，指标本身可能被优化目标反向诱导。

The Show HN project Parsimony is an open benchmark measuring how much extra code AI models add when solving tasks^[11] (69 points). It addresses the decoupling of code volume from value: agents often solve problems by adding a lot of code, and that code carries maintenance, review and blast-radius costs, yet a comparable measure was missing. The design measures change volume as its own dimension rather than only whether tests pass. Worth borrowing is to put net additions alongside correctness in evaluation; the limit is that this is an early project without coverage or task-scope detail, and the metric can be gamed if it becomes an optimisation target.

Show HN 项目 blumify.io/people 用于追踪 AI 领域领袖的公开言论，并观察其观点随时间的变化^[12]（HN 102 分）。它要解决的是观点追踪的成本问题：行业判断往往依赖少数人的表态，而这些表态散落在访谈、播客与社交平台，人工比对前后矛盾非常费力。做法上以人物为轴收集公开言论并做时间线对比。值得借鉴的是把「谁在什么时候说过什么」结构化，是判断行业风向与个人可信度的实用工具；限制是条目未说明信源范围与更新机制，引用断章取义与选择性收录是这类工具的主要风险，结论需要回到原始出处核对。

The Show HN project blumify.io/people tracks public statements by AI leaders and how their views change over time^[12] (102 points). It addresses the cost of tracking positions: industry judgement often rests on a few people's statements, which scatter across interviews, podcasts and social platforms, making manual cross-checking laborious. The design gathers statements per person and compares them on a timeline. Worth borrowing is that structuring who said what and when is a practical way to read industry direction and individual credibility; the limit is that source coverage and update mechanics are unstated, and selective quotation is the main risk.

ZDNet 报道（HN 92 分），荷兰税务部门取消了 Microsoft 365 云迁移计划，转而采用开源方案^[13]。它要解决的是公共部门的数字主权与采购风险管理：把核心办公与数据放在少数外部供应商上，会在合规、成本与议价能力上形成长期依赖，而开源方案把控制权留在机构内部。做法上以「取消迁移 + 转向开源」作为采购决策的修正。对做企业 IT 架构与采购的团队，参考价值是把「可迁移性与退出成本」写进采购评估，比只比单价更重要；限制是报道未给出替代方案的规模与迁移时间表，开源方案的长期运维成本也需要另算。

ZDNet reports, at 92 points on HN, that the Dutch tax authorities cancelled their Microsoft 365 cloud rollout and turned to open source^[13]. It addresses digital sovereignty and procurement risk in the public sector: placing core office work and data with a few external vendors creates long-term dependency in compliance, cost and bargaining power, whereas open source keeps control inside the institution. The decision to cancel and switch is a correction of the procurement path. For enterprise IT architecture and procurement teams the reference is to price portability and exit cost into the evaluation, not just unit cost; the limit is that the replacement's scale and migration timeline are undisclosed.

Show HN 项目 Pulse Local 是一个 Mac 记账应用，其 AI 模型完全在本机运行^[14]（HN 106 分）。它要解决的是个人财务数据的隐私边界：把收支明细交给云端模型意味着最敏感的数据离开设备，而本地推理让小模型承担分类与摘要这类结构化任务已经够用。做法上以本地模型覆盖高频轻量任务，避免上传。值得借鉴的是对数据敏感型应用，「本地模型 + 明确的能力边界」是比云端功能堆叠更好的默认形态；限制是本地模型在复杂语义与多语言票据上的准确率通常不及云端，仓库未给出识别准确率与硬件要求。

The Show HN project Pulse Local is a Mac budgeting app whose AI models run entirely on the machine^[14] (106 points). It addresses the privacy boundary of personal finance data: sending transactions to a cloud model means the most sensitive data leaves the device, whereas local inference already suffices for structured tasks such as categorisation and summarisation. The design covers frequent lightweight tasks locally, avoiding uploads. Worth borrowing is that for data-sensitive apps, local models with explicit capability boundaries beat accumulating cloud features; the limit is that local accuracy on complex or multilingual receipts usually trails cloud models, with no accuracy or hardware requirements given.

路透社报道（经 Techmeme 整理），Jensen Huang 表示 Nvidia 将在五年内提供 2.5 亿美元的加速计算资源，用于资助英国的 AI 安全与 agentic 研究^[22]。它要解决的是安全研究的算力门槛：安全与评测研究往往拿不到前沿算力，导致「谁来验证」这一问题长期依赖模型厂商自证。做法上以算力捐赠换取研究生态，把外部评测能力扶起来。对做安全与评测的团队，参考价值是算力资助会直接决定独立评测能覆盖多深的模型行为，值得跟踪资源分配条件（是否附带披露或审阅条款）；限制是报道未说明申请条件与是否影响研究结论的发布自由，此前已有「评测发表控制权」的争议，需要观察。

Reuters, summarised by Techmeme, reports that Jensen Huang says Nvidia will pledge $250M in accelerated computing resources over five years to fund UK AI safety and agentic research^[22]. It addresses the compute barrier in safety research: safety and evaluation work often cannot obtain frontier compute, leaving verification dependent on the vendors themselves. Donating compute builds an external research ecosystem. For safety and evaluation teams the reference is that compute funding directly determines how deeply independent evaluation can probe model behaviour, so watch the conditions attached (disclosure or review clauses); the limit is that eligibility and any effect on publication freedom are unstated, following earlier disputes over evaluator publication control.

Krebs on Security 报道（经 Techmeme 整理），微软一次性发布更新修补了 1,400 多个安全漏洞，超过 9 月已经创纪录的 972 个——而后者本身已是常态量的约十倍^[23]。它要解决的是漏洞发现流水线的产能问题：当 AI 辅助的代码审计与模糊测试把发现速度推上一个量级，修补与验证就成了新的瓶颈，而企业与运维团队的更新窗口被压缩。做法上厂商以更大批量、更集中方式发布补丁。对做安全运维的团队，参考价值是把「补丁节奏」当成容量规划问题：更新窗口、回归测试与回滚预案都要扩容；限制是报道未说明漏洞的严重性分布与自动化发现所占比例，短期数量激增也不必然意味着实际风险同比例上升。

Krebs on Security, summarised by Techmeme, reports that Microsoft shipped updates fixing more than 1,400 security vulnerabilities, surpassing September's record of 972 — which was itself roughly ten times the usual volume^[23]. It addresses capacity in the vulnerability pipeline: once AI-assisted code audit and fuzzing push discovery up an order of magnitude, patching and verification become the bottleneck while the update window for operations teams shrinks. Vendors respond with larger, more concentrated patch releases. For security operations teams the reference is to treat patch cadence as a capacity-planning problem, scaling update windows, regression testing and rollback plans; the limit is that severity distribution and the share of automated discovery are unstated.

路透社报道（经 Techmeme 整理），ASML 表示在 2027 与 2028 年无法完全满足 AI 驱动的光刻工具需求，同时看到 2029 与 2030 年的订单「强烈兴趣」^[15]。它要解决的是算力供给的物理上限问题：芯片产能取决于光刻设备交付，而设备产能本身受供应链与工艺复杂度限制，因此 AI 算力扩张的节奏最终被上游设备周期约束。做法上公司以订单可见度表达产能紧张。对做算力规划与资本开支的团队，参考价值是把设备交付周期纳入算力路线图，而不是只算芯片设计；限制是公司未给出具体产能数字与被拒绝的需求规模，且这是公司口径，需要与其他设备与代工数据交叉验证。

Reuters, summarised by Techmeme, reports that ASML says it cannot fully meet AI-driven demand for its lithography tools in 2027 and 2028, while seeing “strong interest” in orders for 2029 and 2030^[15]. It addresses the physical ceiling on compute supply: chip capacity depends on lithography tool deliveries, and tool capacity itself is limited by supply chains and process complexity, so the pace of AI compute expansion is ultimately bounded upstream. The company expresses tightness through order visibility. For compute planning and capex teams the reference is to build equipment lead times into the compute roadmap rather than modelling silicon alone; the limit is that no capacity figures or rejected demand are given.

The Verge 报道（经 Techmeme 整理），Meta 把 AI 驱动的「Your Algorithm」功能从 Instagram 扩展到 Facebook 与 Facebook Reels，并表示正在其排序系统中更多地使用大语言模型^[16]。它要解决的是推荐系统的可控性问题：用户长期只能通过行为被动影响推荐，而显式调节把控制权交回用户，同时给平台一个解释推荐结果的接口。做法上用 LLM 参与排序决策，并用自然语言呈现可调参数。对做推荐与内容产品的团队，参考价值是「让用户显式表达偏好」与「用 LLM 改造排序」这两件事可以互相支撑；限制是报道未说明 LLM 在排序中的具体角色与可调项的数量，用户实际获得多少控制权仍待体验验证。

The Verge, summarised by Techmeme, reports that Meta expanded its AI-powered “Your Algorithm” feature from Instagram to Facebook and Facebook Reels, saying it is using LLMs more in its ranking systems^[16]. It addresses controllability in recommendation: users could only influence feeds passively through behaviour, whereas explicit controls return agency and give the platform an interface for explaining results. LLMs participate in ranking decisions while parameters are exposed in natural language. For recommendation and content teams the reference is that letting users state preferences explicitly and using LLMs to reshape ranking can reinforce each other; the limit is that the LLM's specific role and the number of adjustable knobs are unstated.

Epic 创始人 Tim Sweeney 在 X 上表示公司正在进行一次大规模调整，重新分配开发者以「为最大程度的 AI 加速做准备」，并强调目标不是「用同样的方式做同样的事」^[17]（该帖 1.58 万赞）。它要解决的是成熟软件公司如何重排开发方式：引擎与工具链公司如果不把 AI 融入生产流程，会在迭代速度上落后，而全面转型又意味着组织与岗位的重新分配。做法上以管理层公开发声定调，随后进入执行。对做工程组织规划的读者，参考价值是「AI 加速」的落地表现为人员与流程重排，而不是买几个工具；限制是帖文只是方向声明，缺少具体计划、时间表与对现有项目的影响说明，实际效果需要观察后续产品与人员变动。

Epic founder Tim Sweeney said on X that the company is undergoing a large reorganisation, moving developers around to retool for maximum AI acceleration, stressing the aim is not to do the same work the same way^[17] (15.8k likes). It addresses how mature software companies resequence development: an engine and toolchain vendor that does not fold AI into production will fall behind on iteration speed, while a full pivot means reassigning people and roles. Management sets the direction publicly and execution follows. For engineering organisation planning the reference is that AI acceleration shows up as reorganised people and processes, not a few purchased tools; the limit is that the post is a direction statement without plans, timelines or project impact.

HN 上 198 分的条目显示，面向女性与技术从业者的 Grace Hopper Celebration 被取消^[20]。它要解决的是行业多元活动在预算与立场压力下的存续问题：企业赞助是这类活动的财政基础，当行业环境与公司预算变化时，最先被砍的往往是「不直接影响产品」的支出。对做团队文化与人才招聘的读者，参考价值是把多元与人才渠道理解为需要主动投入的长期项目，而不是可随预算波动的营销支出；限制是该条目为活动方公告的转述，取消的具体原因（资金、组织还是外部压力）需要阅读原文确认，不宜作单一归因。

A 198-point HN item indicates that the Grace Hopper Celebration, an event for women and technologists, has been cancelled^[20]. It addresses the survival of diversity-focused industry events under budget and political pressure: corporate sponsorship funds such events, and when industry conditions and budgets shift, spending that does not directly affect products is cut first. For team culture and talent acquisition readers the reference is to treat diversity and talent pipelines as long-term programmes requiring deliberate investment rather than marketing spend that follows budgets; the limit is that the item relays an announcement, so the specific reason needs the original source.

纽约时报报道（经 Techmeme 整理）称，FBI 与 CrowdStrike 合作在得州拆解了 40 多个「笔记本农场」，这些设备被用来帮助朝鲜人员以盗用身份获得美国远程技术岗位^[18]。它要解决的是远程招聘的身份核验缺口：面试与代码测试都可能由他人代做，入职后由被冒名者「持机」维持在线状态，企业直到安全审计或税务异常才发现问题。做法上执法与安全厂商从设备与网络特征入手识别这类农场。对做远程招聘与设备管理的团队，参考价值是把身份核验、设备可信度与网络出口一致性纳入入职流程；限制是报道未给出涉案企业名单与损失规模，企业侧的应对手段（硬件令牌、面签、出口策略）效果也缺少公开数据。

The New York Times, summarised by Techmeme, reports that the FBI, working with CrowdStrike, dismantled 40+ “laptop farms” across Texas used to help North Koreans obtain remote US tech jobs under stolen identities^[18]. It addresses the identity-verification gap in remote hiring: interviews and coding tests can be done by someone else, and after hiring the equipment is kept online by a stand-in, with companies discovering the problem only in a security audit or tax anomaly. Law enforcement and security vendors identify such farms from device and network characteristics. For remote hiring and device management teams the reference is to bring identity verification, device trust and network-egress consistency into onboarding; the limit is that no affected companies or loss figures are given.