Special Edition

Agents Get Permissions, AI Hardware Booms and Governance Goes Global

A Special Edition round-up: GPT-6 Cyber, Tencent Hy4 and Claude Personal Finance, the Marvell–Lumentum deal and AMD's data-centre surge, METR's 1.9x finding, and new AI policy in China and the UK.

By H.AI Tools Team4 stories
Products in this issue: ChatGPT, Claude, Gemini, ChatExcel, Excelly-AI
In this issue

The short version

  • Three launches gave AI access to sensitive work: GPT-6 Cyber for vetted security teams, Tencent Hy4 for operating computers, and Claude Personal Finance for read-only bank data.
  • Marvell agreed to buy Lumentum for an implied $115 billion, and AMD's data-centre revenue nearly doubled to $8.5 billion. Connectivity is now as strategic as chips.
  • METR estimates Anthropic's model progress has sped up about 1.9 times, with wide uncertainty and some evidence withheld.
  • China ratified the agreement creating a World AI Cooperation Organization based in Shanghai, and the UK backed Potential Sciences with investment and national supercomputer time.

This Special Edition covers the week of November 1 to 7. The common thread is trust. AI systems were given access to security research, desktop software and bank accounts. The hardware behind them attracted record deals. Independent evaluators and governments tried to keep up. The sections below connect the week's stories and turn each into something you can check or apply.

01

Agents get permissions: security research, computer use and bank data

Three launches with different levels of access

OpenAI released GPT-6 Cyber on November 5 for finding, verifying and proposing fixes for software vulnerabilities. OpenAI rates its cyber capability "Critical", the highest level in its Preparedness Framework. Access is limited to approved organisations in its Daybreak Red programme, with identity checks, usage restrictions and monitoring. Tencent (腾讯) released Hy4, the first model in its Hunyuan (混元) family that can understand images in depth and operate a computer through both the graphical interface and the command line. Tencent reports strong scores on the OSWorld computer-use benchmark; these are company figures, and the weights had not been released at launch.

Anthropic began rolling out Claude Personal Finance to US subscribers on Pro, Max and Ultra plans. It connects bank and card accounts through Plaid, so Claude never sees your bank password. It reads balances and transactions and can analyse spending, but it cannot move money or make investments. Conversations in the dedicated finance area are excluded from model training.

A permissions ladder for your own agents

The three launches sit on different rungs of the same ladder: read-only access, actions with approval, and actions without approval. Personal Finance stays on the first rung. GPT-6 Cyber can act, but only for vetted teams under monitoring. Computer-use models like Hy4 can act across any application they can see, which is powerful and needs the most care.

Before connecting an assistant such as ChatGPT or Claude to your accounts, decide which rung each connection deserves. Read-only is enough for analysis, such as "summarise last month's spending by category". Approval should be required for anything that sends, buys, deletes or changes settings. Reserve unattended action for low-risk, reversible tasks in a test account. Review connected apps once a month and remove any you no longer use. Our ChatGPT vs Claude guide compares the two assistants' broader features.

Get HAI Daily in your inbox

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

02

AI hardware: connectivity deals and a data-centre surge

What was announced

Marvell agreed to combine with Lumentum on November 2 in an all-stock deal with an implied enterprise value of about $115 billion. The aim is to pair Marvell's chips with Lumentum's optical technology, which moves data between processors in AI data centres. The deal is expected to close in the first half of 2028, subject to shareholder and regulatory approval. On November 3, AMD reported third-quarter revenue of $13.4 billion, up 45%, with data-centre revenue up 97% to a record $8.5 billion. It credited demand for EPYC server processors and the launch of its Instinct MI450 accelerators.

Later in the week, OpenAI was reported to have closed the first $11 billion of a new funding round, covered in our November 9 issue.

Read an earnings release with AI, then check it

Earnings releases are where AI demand shows up in hard numbers, and an AI assistant can help you read them quickly if you check its work. Download the company's release or quarterly filing, then ask a spreadsheet-capable assistant to build a table of revenue by segment for this quarter and the same quarter last year, with growth rates.

ChatExcel can work directly on a table you upload, and Excelly-AI can write the growth-rate formulas for your own sheet. Then verify. Recalculate one growth rate yourself, confirm each figure against the page it came from, and label any number that came from commentary rather than the filing. With AMD, for example, the 97% data-centre growth is in the company's own release. Our Excelly-AI review and the Data Analysis category list more tools. This is a reading exercise, not investment advice.

03

Measuring AI progress: METR's 1.9x estimate

What METR found

METR, an independent AI evaluation organisation, published an assessment on November 5 of how fast Anthropic's models are improving. Using Anthropic's internal capabilities index, METR estimates that the rate of progress since October 2025 is about 1.9 times the earlier pace, with a 95% confidence range of 1.5 to 2.8 times. Its rough judgement is that around 80% of the speed-up comes from using AI in research and development rather than from more computing power. METR stressed the uncertainty and noted that Anthropic redacted some of the evidence.

Build a small evaluation set for your own work

Independent assessments matter because vendor benchmarks rarely match your work. If models are improving faster, the tool that was best for your task three months ago may not be best today. The practical response is a small, reusable test set.

Collect ten real tasks from your work with known good answers: three easy, four typical and three hard. Save the prompts and the expected results. Each time a new model appears in Claude, Gemini or ChatGPT, run the set and record how many it gets right, how long it takes and what it costs. Twenty minutes per model gives you better evidence than any leaderboard. Keep sensitive data out of the set, or use anonymised versions.

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

04

Policy: a global AI body in Shanghai and UK compute for chemistry

China and the UK

On November 6, the Standing Committee of China's National People's Congress ratified the agreement establishing the World AI Cooperation Organization. China's government signed it in Shanghai on July 16, and the organisation will have its headquarters in Shanghai to promote international cooperation and governance on AI.

In the UK, the government's Sovereign AI unit backed Potential Sciences, a start-up using physics-based AI to design molecules for drug discovery, as part of an eight-figure funding round. The support includes access to computing through the AI Research Resource, including the Isambard-AI supercomputer in Bristol. The company says its approach generates its own training data from physics-based calculations rather than relying on scarce experimental data. That claim has not yet been tested in public.

What to take from it

The two stories show different routes to the same goal. China is building international institutions around AI governance. The UK is pairing public investment with national computing capacity for specific companies. For researchers and start-ups, the useful lesson is that public compute programmes exist and can be applied for. Check the official programme pages in your own country for eligibility and deadlines before assuming that only private cloud capacity is available.

What to try next

What to watch next

Watch for wider access to GPT-6 Cyber beyond the first approved teams, a cheaper Hy4 Flash version reported to be on the way, and whether other assistants follow Claude into read-only financial data. The Marvell–Lumentum deal now enters a long approval process. For your own work, the lesson of the week is to give agents the lowest level of access that does the job and to test new models on your own tasks.

Explore Chatbots, Data Analysis 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