八月二十七日

Anthropic made Claude in Chrome generally available on every paid Claude plan and introduced a separate browser built into Claude Cowork. The Chrome extension works inside a user's existing browser and can use current logins to read pages, click, type, navigate, and complete forms. Cowork's browser instead opens in the desktop app's side panel, stays separate from personal tabs, bookmarks, and passwords, and is designed for delegated tasks such as gathering research or collecting invoices from a portal. Users can import selected site logins, with banking, email, and single sign-on excluded unless explicitly included. The built-in browser is rolling out to Pro, Max, and Team plans on macOS, Windows, and Linux, while Enterprise admins can enable it now.

Anthropic 已向所有 Claude 付费方案正式开放 Claude in Chrome，并在 Claude Cowork 中加入一套独立的内置 browser。Chrome extension 在用户现有浏览器中工作，可以沿用当前登录状态，读取页面、点击、输入、跳转并完成表单。Cowork browser 则在 desktop app 的 side panel 中打开，与个人 tabs、bookmarks 和 passwords 隔离，更适合委托式任务，例如收集研究资料或从 vendor portal 下载 invoices。用户可以按站点导入指定登录信息，banking、email 和 single sign-on 默认不会被带入，除非用户明确选择。内置 browser 正在向 macOS、Windows 和 Linux 上的 Pro、Max 与 Team 方案逐步推出，Enterprise admins 现在即可启用。

Claude in Chrome can now approve actions it determines are safe instead of asking the user every time. Anthropic says web content is screened for likely prompt injection before Claude acts, while a separate classifier checks proposed actions against the user's original request and blocks mismatches. On a stronger evaluation using attacks from professional red-teamers, attacks reaching the unsafeguarded models succeeded 17.6% of the time against Opus 4.5 and 3.8% against Opus 5. With probes and the automatic-approval safety classifier, Anthropic reported no successful attacks against Sonnet 5 or Opus 5 and a 0.3% success rate against Fable 5, all manually verified as low severity. Anthropic still calls prompt injection a moving target and says the safeguards reduce rather than eliminate risk.

Claude in Chrome 现在可以自动批准其判断为安全的 actions，不再要求用户逐次确认。Anthropic 表示，web content 在 Claude 执行前会先经过 probe 检查，以发现可能的 prompt injection；另一个 classifier 则会把拟执行 action 与用户最初的请求进行比对，并阻止不一致的操作。在一项使用专业 red-teamers 攻击样本的更强评测中，攻击在没有额外 safeguards 时一旦到达 model，对 Opus 4.5 和 Opus 5 的成功率分别为 17.6% 与 3.8%。加入 probes 和 automatic-approval safety classifier 后，Anthropic 报告 Sonnet 5 与 Opus 5 没有成功攻击，Fable 5 的成功率为 0.3%，且人工复核确认都属于低严重度场景。Anthropic 仍强调 prompt injection 是持续变化的攻击面，这些 safeguards 只能降低风险，无法彻底消除风险。

Box CEO Aaron Levie tied the company's Q2 revenue of $321.1 million, up 9% or 11% in constant currency, and its raised $1.290 billion full-year revenue target to enterprise demand for AI-ready content. His broader argument is that capable models remain limited without secure access to contracts, research, financial documents, marketing assets, product roadmaps, and other unstructured corporate knowledge. Enterprises therefore need an applied AI layer that can connect agents to context, swap models as cost and performance change, and enforce guardrails, audit logs, and real-time security alerts. Levie sees the durable opportunity in bringing intelligence to governed workflows without forcing companies to migrate their data into a separate system.

Box CEO Aaron Levie 将公司 Q2 revenue 达到 3.211 亿美元，同比增长 9%、按 constant currency 计算增长 11%，以及将全年 revenue target 上调至 12.90 亿美元，与企业对 AI-ready content 的需求联系起来。他更广泛的判断是，即使 model 能力很强，如果不能安全访问 contracts、research、financial documents、marketing assets、product roadmaps 等非结构化企业知识，其价值依然受限。因此，企业需要一个 applied AI layer，让 agents 能连接正确 context，随着成本和性能变化切换 models，并通过 guardrails、audit logs 与 real-time security alerts 实施治理。Levie 认为，持久的机会在于不要求企业把 data 迁移到另一个系统，就能把 intelligence 带入受治理的 workflows。

Vercel CEO Guillermo Rauch highlighted multi-region compute with failover, default capacity for as many as 10,000 concurrent sandboxes, and a ramp rate of 5,000 vCPUs per minute, with more regions planned. He also previewed a security dashboard and the `vercel security check` CLI, which can enlist agents to improve a project's security posture either with a human in the loop or on a scheduled run. Together, the releases position agent infrastructure around two production constraints: rapid, globally distributed execution and security checks that can run through the same automation surface as the rest of the development workflow.

Vercel CEO Guillermo Rauch 介绍了支持 failover 的 multi-region compute、默认最多 10,000 个 concurrent sandboxes，以及每分钟扩展 5,000 vCPUs 的能力，并表示后续还会增加更多 regions。他还预告了一套 security dashboard 和 `vercel security check` CLI，可以让 agents 在 human-in-the-loop 模式下改善项目的 security posture，也可以按计划自动运行。这两项发布共同指向 agent infrastructure 的两项 production 约束：快速、全球分布式的执行能力，以及能够通过开发 workflow 同一套 automation surface 持续运行的安全检查。

SPC partner Aditya Agarwal announced DeepCogito's $43 million Series A and argued that the AI frontier will be determined by post-training. He described the lab's focus as large-scale reinforcement learning and recursive self-improvement through iterated distillation and amplification, or IDA. According to Agarwal, DeepCogito has already demonstrated the method publicly on models ranging from 3 billion to more than 600 billion parameters.

SPC partner Aditya Agarwal 宣布 DeepCogito 完成 4,300 万美元 Series A，并提出 AI frontier 将由 post-training 决定。他将这家 research lab 的重点概括为通过 iterated distillation and amplification，也就是 IDA，开展 large-scale reinforcement learning 与 recursive self-improvement。Agarwal 表示，DeepCogito 已在 30 亿到超过 6,000 亿 parameters 的 models 上公开验证这套方法。

Claude Code builder Thariq introduced a `SendFeedback` tool that lets users ask Claude to draft a product report and then approve it, replacing the need to invoke `/feedback` and write the report manually. The small interaction change applies the product's central idea to its own improvement loop: Claude can assemble the context for a report, while the user retains the final approval step.

Claude Code builder Thariq 发布了 `SendFeedback` tool。用户现在可以让 Claude 起草 product report，再由自己批准，不必调用 `/feedback` 后手动撰写。这项看似很小的交互变化，把产品的核心理念也用于自身 improvement loop：Claude 负责整理 report 所需的 context，用户保留最终批准环节。

The Takeaway: Walleye's owner-operator Will is treating AI fluency as mandatory operating leverage for a roughly 400-person hedge fund, pairing executive pressure with training, shared experimentation, internal data, and tools embedded in investment work. His sharpest cultural rule is: “Using ChatGPT is not cheating.” Walleye began its internal AI program two years earlier after an analyst demonstrated a tool designed to make his own research work more efficient. That effort became Current, an internal system used by every fundamental stock-picking team and especially heavily during earnings periods.

核心结论： Walleye 的 owner-operator Will 把 AI fluency 视为一家约 400 人 hedge fund 的必备 operating leverage，并将管理层推动、全员培训、共享实验、内部 data 和嵌入投资工作的 tools 结合起来。他最鲜明的文化原则是：“Using ChatGPT is not cheating.” Walleye 在两年前启动内部 AI program，起点是一名 analyst 展示了一个旨在提升自身 research 效率的 tool。这个项目后来发展为 Current，供所有 fundamental stock-picking teams 使用，在 earnings periods 尤其活跃。

The rollout is organizational, not just technical. Will said about 75% of the firm actively uses ChatGPT or a similar product every week, almost every day, and roughly one third uses AI coding tools. Mandatory baseline training extends across investment, accounting, finance, compliance, and legal teams. Weekly meetups, usage leaderboards, incentives for recommending deployable tools, and visible executive participation make experimentation social. The goal is not to hide imperfect demos but to normalize learning before the tools become perfect.

这套 rollout 不只是技术项目，更是组织变革。Will 表示，约 75% 的员工每周都会主动使用 ChatGPT 或类似产品，几乎达到每天使用；大约三分之一的员工使用 AI coding tools。强制性的基础培训覆盖 investment、accounting、finance、compliance 与 legal teams。每周 meetups、usage leaderboards、推荐可部署 tools 的激励机制，以及管理层亲自参与，让实验成为一种组织内的社交活动。目标不是隐藏不完美的 demos，而是在 tools 变得完美之前，先让持续学习成为常态。

The playbook depends on proprietary context. Walleye records most internal calls, processes transcripts, and is building a data strategy that can connect qualitative information with market, accounting, and other numerical data. Current already combines analyst notes, broker material, PDFs, earnings transcripts, and stock-relevant information. Will said drafting a memo from his own bullet points and prior writing can reduce a four-to-five-hour task to about 15 minutes, freeing attention for higher-level decisions. His longer-term view still leaves room for human investors where the situation is novel, the precedents are sparse, and judgment must operate in a “fuzzy mess.”

这套 playbook 的基础是 proprietary context。Walleye 会记录大多数内部 calls、处理 transcripts，并建立能够把 qualitative information 与 market、accounting 等 numerical data 连接起来的 data strategy。Current 已经汇集 analyst notes、broker materials、PDFs、earnings transcripts 与其他和 stocks 有关的信息。Will 表示，根据自己的 bullet points 与历史写作起草 memo，可以把原本四到五小时的任务缩短到约 15 分钟，把注意力释放给更高层次的决策。他的长期判断仍然为 human investors 留出空间，尤其是在情境新颖、可参考先例很少、判断必须面对“fuzzy mess”的时候。