Special Edition

Proactive Agents, Big Chip Orders and AI Security Gets Serious

A Special Edition round-up: Claude Duet and OpenAI on Google Cloud, China's new national data group, Nvidia chip orders and DeepSeek's funding, the AI-found PHP2Shell flaws, OpenAI's offline training, and new AI policy in China and the UK.

By H.AI Tools Team4 stories
Products in this issue: Claude, ChatGPT, Kimi, DeepSeek, Snyk, GitHub Copilot
In this issue

The short version

  • Claude Duet opened as a research preview for US Pro and Max users: an assistant that checks your connected apps and starts work without being asked, but still needs approval to send or spend.
  • OpenAI's models arrived on Google Cloud in private preview, so all three major clouds now sell them.
  • Chinese tech giants reportedly ordered over 3 million Nvidia RTX PRO 5500 cards, and DeepSeek was reported to be closing a round of about ¥100 billion.
  • AI helped find the critical PHP2Shell flaws, OpenAI moved frontier training off the live internet, and both China and the UK set out new AI policies on jobs and security.

This Special Edition covers October 19 to 24. Assistants started to act without being asked, the companies building AI placed very large orders for chips and capital, and security became a practical concern rather than a theoretical one. Governments in China and the UK responded with plans for jobs and security. The sections below connect the stories and suggest what to do with each.

01

Assistants that start the work

What was launched

Anthropic opened a research preview of Claude Duet on October 21. It is an always-on assistant that checks your connected apps every few hours, picks up to-dos, sends a morning briefing and keeps working on tasks in the background. It is available by waitlist to US subscribers on Pro and Max plans, as a standalone app and on claude.ai, at no extra cost during the preview but within normal usage limits. Actions that send messages, spend money, sign or delete things still need your approval. Anthropic warns that it can misread a thread or misjudge what is urgent.

The same day, Google Cloud began a private preview of OpenAI models for selected enterprise customers: GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna, with support for Codex workflows. Customers can buy them through the cloud marketplace and count the spend against existing Google Cloud commitments. Prices, zero data retention and some security options were still to come. Moonshot (月之暗面) also launched Kimi K3.1 Pro on October 23 and promised open weights, which arrived in early November, as covered in our November 5 issue.

Setting up a proactive assistant safely

An assistant that acts first is useful only if you trust what it sees and what it can do. Start with read-only connections, such as calendar and email, and let it produce briefings for a week before granting any action. Check each briefing against your inbox: did it miss anything important or flag noise? Then allow drafting, but keep sending and spending behind approval permanently.

For companies, OpenAI's arrival on Google Cloud means model choice no longer has to follow your cloud provider. If you already have a cloud commitment, ask your account team which models count against it before signing separate API contracts. Claude and ChatGPT both offer business plans with admin controls; our Claude review covers its plans.

Get HAI Daily in your inbox

The day's AI news, what it changes, and which tools to try.

02

Data, chips and capital

What was announced and reported

China established the National Data Group in Xiong'an, Hebei, on October 20, a central state-owned company whose main business is developing and using data resources. It complements the National Data Administration: the administration sets policy, while the group runs market operations and infrastructure.

Reuters reported on October 20, citing people familiar with the matter, that ByteDance (字节跳动), Alibaba (阿里巴巴) and Tencent (腾讯) had together ordered more than 3 million Nvidia RTX PRO 5500 cards, worth over $35 billion at list price, mainly for AI inference. Demand exceeds planned production, and Nvidia reportedly cannot guarantee full delivery. The companies did not comment.

Reuters also reported on October 22 that DeepSeek (深度求索) was wrapping up a funding round of about ¥100 billion, around $15 billion, with several Chinese technology and industrial companies taking part. The amount and investor list could still change, and DeepSeek did not comment.

What this means for people using Chinese AI services

Large chip orders and fresh capital usually mean more capacity and lower prices for users of Chinese models over the next year, though supply limits may slow that. If you rely on DeepSeek, Kimi or other Chinese assistants, track their price lists and usage limits each quarter, and keep notes on where your data is processed. Our DeepSeek review covers its current plans.

03

AI security in practice: PHP2Shell and offline training

What happened

On October 21, PHP released emergency patches for two critical vulnerabilities known together as PHP2Shell. Researchers at security start-up Calif, using an AI system, and at Wiz found that the flaws could let attackers run code on sites built with WordPress, Drupal and Joomla without logging in. The fixed versions are PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35. The UK's National Cyber Security Centre urged organisations to patch immediately and to take systems offline if they cannot apply mitigations in time.

On October 23, OpenAI said it now trains its frontier models without access to the live internet, after incidents in which models briefly reached the internet through misconfigurations during training. Models now practise browsing in an offline environment built from cached web pages, and OpenAI expanded its red-team testing. Training has resumed under the new rules.

What to do this week

If you run any website on PHP, check your version today and update to a patched release; ask your hosting provider if you are not sure. Then set up a routine so the next critical patch does not depend on reading the news: subscribe to your platform's security announcements and run a scanner such as Snyk on your code and dependencies. Coding assistants such as GitHub Copilot can help apply and test upgrades, but review every change.

The PHP2Shell case shows AI finding serious flaws faster than before. Expect more emergency patches, and shorten the time between a fix being released and your systems running it.

Not sure which tool fits your task?Browse the H.AI Tools directory by category, then compare a shortlist.

04

Policy: AI and jobs in China, AI security in the UK

What was announced

China's State Council General Office published guidelines on October 23 on adapting to AI while protecting employment. They set out 18 tasks in six areas, including creating new AI-related jobs, slowing the displacement of workers, a risk assessment mechanism for the employment effects of AI, special AI training programmes, stronger social protection for flexible and platform workers, and employment monitoring.

On October 24, the UK government set out about £600 million of AI security funding. It includes £165 million for a National AI Biosecurity Initiative, about £100 million of which funds a new AI Biosecurity Lab, an extra £120 million for the AI Security Institute and money for cyber defence of critical infrastructure. AI companies are to provide model access and computing credits on top.

What to take from it

Both governments are treating AI's effects on people as a policy task. For employers, the Chinese guidelines point toward assessing employment effects before replacing roles with AI, and offering retraining. That is good practice anywhere. Before automating a task, list the people whose work changes, what new skills they need and how you will train them. For security teams in the UK, the new funding signals more official guidance and testing to come; follow the AI Security Institute's publications.

What to try next

What to watch next

Watch for wider access to Claude Duet, general availability and pricing of OpenAI models on Google Cloud, and whether the reported chip orders are delivered in full. For your own work, give proactive assistants read-only access first, patch PHP sites now and plan training before you automate jobs.

Explore Chatbots, AI Security and Code Generation to build your shortlist.

How we put this issue together

The H.AI Tools Team checks each story against the official sources linked in the text and removes claims we could not confirm. Product links lead to H.AI Tools directory profiles. Unless an issue says otherwise, product notes describe vendor documentation, not our own benchmark tests.

Go deeper