UK AI Cyberattack Benchmarks: Autonomous Hacking Accelerates

Samantha: Welcome to The State of Tech, The European Edition, Friday May fifteen, 2026. I'm Samantha Lawrence.

Bob: And I'm Bob Russell. Today: Nvidia's high-stakes China chip mission, a six-point-six billion dollar payday for OpenAI staff, China's quantum leap with Jiuzhang 4.0, the UK measuring how fast AI can hack on its own, Gemini taking over your Android keyboard, and WhatsApp's new incognito chat with Meta AI. Let's start with Nvidia in Beijing.

Nvidia's Jensen Huang lands in Beijing chasing a breakthrough as H200 chip shipments to China stall.

Bob: Jensen Huang is in Beijing this week, and he's not there for the food. Washington has cleared roughly ten Chinese companies, names like Alibaba, Tencent and ByteDance, to buy Nvidia's H200, the AI chip everyone wants. But here's the catch: not a single one has actually been delivered yet.

Samantha: So the licences are on paper, the chips are not on planes. Huang travelled as part of a US delegation led by President Trump, which tells you how political this has become. Nvidia desperately wants the China revenue back. Beijing wants the compute. And Washington wants leverage without handing over the crown jewels.

Bob: And the H200 isn't Nvidia's absolute top chip, it's the cut-down export version. But it's still powerful enough to train serious AI models. That's why Anthropic published a paper this week arguing that whoever has the most compute, basically wins the AI race. It's a strategic resource now, like oil in the seventies.

Samantha: The deeper issue is that nothing about this deal is stable. One executive order can flip it. Chinese buyers know that, so even if shipments start, they'll keep pouring money into Huawei's homegrown alternatives. Nvidia risks training its own competitor by being too slow.

Bob: For Europe this matters more than people think. We're not in the export-control driver's seat, we're a passenger. European AI labs, Mistral, Aleph Alpha, the research institutes, they all buy the same Nvidia chips on the same global queue. If Washington reroutes supply toward or away from China, European allocations move too.

Samantha: And there's no European H200. The EU Chips Act is real, but it's targeting mature nodes, the chips in your car and your washing machine, not bleeding-edge AI accelerators. Which means Europe's AI ambitions are, for the foreseeable future, hostage to a negotiation happening in a Beijing hotel between Huang and Chinese regulators.

Bob: The honest read: even if Huang gets a handshake this week, the underlying tension doesn't go away. Export controls have become the main instrument of tech policy, and chips are the new currency. Expect more of these dramatic trips, and more delays.

---

OpenAI employees set to cash out six-point-six billion dollars in a secondary share sale.

Samantha: Onto a very different kind of chip count, dollars. OpenAI employees are reportedly about to share a six-point-six billion dollar payday through a so-called secondary sale. That means staff sell existing shares to new investors, the company itself doesn't go public, but the valuation keeps climbing.

Bob: And it's a smart move for OpenAI as an employer. The AI talent war is brutal. Anthropic, Google DeepMind, Meta, they're all dangling huge packages. By letting your engineers actually convert paper wealth into a house or a pension, you give them a reason to stay rather than jump ship.

Samantha: But you have to ask, who's buying? Investors are stepping in at a valuation that, last we heard, was deep into the hundreds of billions. For that to make sense, OpenAI needs to roughly own the future of enterprise AI. Any serious slip, technical, regulatory, or competitive, and the maths gets ugly fast.

Bob: And this is where the question of whether the AI sector is still being priced sensibly comes in. We're seeing valuations that, in any other cycle, would trigger alarm bells. The argument from believers is that AI is genuinely different, the productivity gains are real. The argument from sceptics is, that's what people said in 1999.

Samantha: For European pension funds and sovereign wealth funds, some of which are reportedly in these rounds via intermediaries, that's the uncomfortable bit. European savers may be quietly exposed to the most concentrated bet in tech history, without ever clicking buy.

Bob: And for European AI startups, this sets the salary bar at a level they simply cannot match. If your best machine-learning engineer can fly to San Francisco and cash a life-changing cheque within two years, retention in Paris or Munich becomes very hard.

Samantha: Bottom line: OpenAI gets to keep its team, investors double down, and the rest of us find out later whether the price tag was visionary or vertigo.

Samantha: Quick interruption. If you listen to The State of Tech regularly, hit that like button and subscribe, that way you'll never miss an episode. Okay, moving on.

---

China's Jiuzhang 4.0 quantum machine claims a calculation ten to the fifty-fourth times faster than top supercomputers.

Bob: Chinese scientists at the University of Science and Technology of China have unveiled Jiuzhang 4.0, a quantum computing prototype that uses light particles, photons, to do calculations. They ran a very specific maths problem and claim a speed-up of ten to the power fifty-four over the world's best classical supercomputer. That number is so large it stops being meaningful.

Samantha: Important caveat: this is one narrow benchmark, called Gaussian boson sampling. It's not the same as running your spreadsheet faster or cracking encryption tomorrow. But the engineering jump is real, they're manipulating up to three thousand photons at once. The previous version handled far fewer.

Bob: And it's a clear signal in the quantum race. The US has IBM and Google focused on a different approach using superconducting circuits. China is leaning hard into photonics, light-based quantum machines. Both have trade-offs, but China just planted a very visible flag.

Samantha: For Europe, this is a wake-up moment. The EU has its Quantum Flagship programme, and countries like the Netherlands, France and Germany have serious research groups, QuTech in Delft, for instance. But funding compared to China and the US is modest, and we don't have a national champion of this scale.

Bob: The longer-term concern is encryption. Today's banking, government, and medical data is protected by maths that a sufficiently powerful quantum computer could one day unpick. Jiuzhang 4.0 doesn't do that, but every milestone shortens the runway. European regulators are quietly pushing post-quantum cryptography, and stories like this one explain why.

Samantha: So no, your laptop isn't obsolete this morning. But the gap between physics demo and security threat is closing faster than the policy world is moving.

---

UK AI Safety Institute warns autonomous AI cyberattack ability is improving faster than expected.

Samantha: The UK AI Safety Institute, a government body that tests how dangerous AI systems are getting, just published research on autonomous cyberattacks. The plain question they asked: can an AI, with no human hacker watching, find and exploit security holes in real software, and how fast is that getting better?

Bob: Their answer, in short: faster than people thought. The institute is deliberately building a yardstick, so governments and companies can see the trend line before it's too late. The numbers show clear jumps with each new generation of models on tasks like reconnaissance, finding vulnerabilities, and chaining steps into an actual exploit.

Samantha: What's striking is who's doing this measurement. It's not the AI makers themselves grading their own homework, it's a national body. That matters because if you let labs publish only the capabilities they're comfortable with, you get a sanitised picture. AISI is essentially saying, we'll be the referee.

Bob: And the asymmetry is the scary part. Attackers only need one working exploit. Defenders need to patch everything. If AI cheapens the cost of finding flaws, attackers get the productivity gain first, because they don't need approvals, compliance checks, or change-management meetings.

Samantha: For European companies and governments, this is urgent. Critical infrastructure, hospitals, energy grids, town councils, has been pummelled by ransomware for years already, often with sloppy attackers. Make those attackers smarter and cheaper, and the threat scales.

Bob: The European equivalent is taking shape. The EU AI Office in Brussels handles general-purpose AI under the AI Act, and ENISA covers cybersecurity. But neither is yet doing the specific capability evaluations AISI is publishing. There's quiet talk of an EU AI Safety hub modelled on the UK approach.

Samantha: The takeaway for businesses: assume the attacker on the other side has an AI assistant. Treat AI-assisted defence as a now-budget item, not a 2027 project.

---

Bob: Alright, to close out, two things you can actually use today. First one, an AI feature most Android users are about to get for free, whether they asked or not.

Google is rebuilding Gboard's microphone button with Gemini, turning your phone into a much smarter dictation tool.

Bob: Google is plugging its Gemini AI into Gboard, the default keyboard on Android phones. From now on, when you tap the little microphone icon to speak instead of type, it's Gemini converting your voice to text. And it doesn't just transcribe, it cleans up the rambling, adds punctuation, and can rephrase what you said into a tidier sentence.

Samantha: Which sounds harmless until you remember an entire category of startups built businesses on exactly this. Otter, smaller players doing meeting notes and dictation, paid subscriptions starting at around ten euros a month. If the basic experience on every Android phone is suddenly this good, the reason to install a separate app shrinks dramatically.

Bob: It's the classic platform move. A feature gets validated by startups for years, the platform owner watches, and then absorbs it into the operating system. Sherlocking, as the saying goes. If you're building an AI product on top of someone else's distribution, this is your nightmare scenario in a single news item.

Samantha: Practical bit for listeners. The rollout has started on Pixel phones first, with broader Android availability coming in waves. In Europe, expect a slightly slower arrival because of how Google handles Gemini features under the Digital Markets Act, certain integrations need a tick-box from Brussels first.

Bob: To try it: open Gboard on a recent Android phone, tap the microphone, and start talking naturally. The clean-up happens automatically. You can dictate an email, a WhatsApp message, even a longer note. It's genuinely useful, especially in the car or while walking.

Samantha: One thing to keep in mind, dictation goes through Google's servers for the AI tidy-up. If you're handling sensitive client information, that's worth knowing. For everyday messages, it's a real upgrade and it costs you nothing.

Bob: A free upgrade, with the small footnote that one more daily moment of your life now flows through a Google AI.

---

WhatsApp launches Incognito Chat with Meta AI, promising no history and no training data.

Samantha: And second find, also free, also worth a try, but with a bigger asterisk. WhatsApp is rolling out a feature called Incognito Chat. It lets you talk to Meta's AI assistant without the conversation being tied to your WhatsApp account. Meta says these chats aren't stored, don't appear in your chat list, and aren't used to train the AI further.

Bob: The reason Meta is launching this now is obvious. ChatGPT, Claude, Gemini, none of them are tied to your social identity. People ask AI chatbots deeply personal questions, medical, financial, relationship stuff, and they don't want that linked to the account where they message their mother.

Samantha: But, a privacy promise from Meta should not be taken at face value. The fine print matters. Even in incognito mode, metadata, the fact that you used the feature, when, from which device, can still be logged for what they call abuse prevention. And Meta hasn't confirmed exactly how the European rollout interacts with GDPR and the Digital Markets Act.

Bob: It is also not the same as Signal or a model running locally on your laptop. With Signal, the company genuinely cannot read your messages. With a local AI, nothing leaves your device. Incognito Chat is Meta saying trust us, which is a different category.

Samantha: To try it: open WhatsApp, start a new chat, and look for the Meta AI option with an Incognito toggle. Availability is rolling out, the US first, EU users may see a delayed launch while Brussels reviews the privacy terms.

Bob: If you want a useful test, ask it the kind of thing you'd never type into a regular search box. A health worry, a salary question, a tricky family situation. Then decide for yourself whether Meta has earned the trust.

Samantha: Today we covered: Nvidia's China chip mission, OpenAI's six-point-six billion payday, China's Jiuzhang quantum breakthrough, the UK measuring AI cyberattack speed, Gemini taking over Gboard dictation, and WhatsApp's Incognito Chat with Meta AI.

Bob: Want to know more or react? Visit stateoftech.eu or email us at info@doorzetters.net.

Bob: State of Tech, the tech world in 15 minutes.