Anthropic Mythos, AI Agents & Cybersecurity, China Digital Humans, Nvidia Chip Restrictions

Samantha: Welcome to The State of Tech, The European Edition, Sunday April nineteen, 2026. I'm Samantha Lawrence.

Bob: And I'm Bob Russell. Today: Barclays sounds the alarm on Anthropic's Mythos AI model, the explosion of AI agents and the cybersecurity headache they bring, China cracks down on digital humans, Nvidia's Jensen Huang pushes back on chip restrictions, the AI-meets-sustainability market heading for nearly a trillion dollars, and DARPA's wild plan to mix and match quantum computers. Let's start with the story everyone in banking is talking about this weekend.

A bank CEO publicly warns the world about an AI model.

Samantha: Bob, when the boss of Barclays goes on the record calling an AI model a serious threat to the global banking system, people listen. That's exactly what happened with CS Venkatakrishnan and Anthropic's new model, Mythos.

Bob: And this isn't vague hand-waving. Mythos is currently in a gated preview under something called Project Glasswing. Officially, it's being used to find vulnerabilities. Unofficially, the concern is that the same model, with its advanced coding skills and agentic abilities, could launch cyberattacks at a level we haven't really seen before.

Samantha: And the political backdrop is just as interesting. The White House met with Anthropic this week. Remember, there was a dispute with the Pentagon not long ago, so this feels like a reset.

Bob: It does. And on top of that, the US is planning to deploy Mythos inside federal agencies. So you have this strange split: on one hand, governments warning the private sector to be careful, and on the other hand, governments themselves rolling the same technology out internally.

Samantha: Officials in the US, Canada and Britain are all now in active conversations with bank executives. Banking is an obvious target — high value, lots of interconnected systems, and regulators breathing down everyone's neck.

Bob: The asymmetry here is what worries me. A well-defended bank has thousands of people working on security. An attacker with an agentic model like Mythos could, in theory, probe systems around the clock, write custom exploits, and chain together attacks without a human in the loop. That changes the economics of cybercrime.

Samantha: And the bottom line is that cybersecurity used to be a race between humans and humans. Now it's humans plus AI, and very soon, AI versus AI.

Bob: For Europe, this lands right in the middle of the AI Act implementation. British banks are already in the conversation because the UK Treasury is one of the three governments involved. But the EU will be watching this closely too, because critical infrastructure rules under NIS2 and the AI Act's high-risk provisions will have to contend with models like Mythos.

Samantha: European banks — think Deutsche Bank, BNP Paribas, ING — they all use essentially the same technology stacks and face the same exposure. So if the US and UK are briefing their banks, you can bet Frankfurt, Paris and Amsterdam are getting similar phone calls.

Bob: And there's a policy question here too. If Washington is happy to deploy Mythos across its federal agencies, will European governments want access to the same capability, or will they want an EU alternative? Digital sovereignty isn't just a slogan anymore — it's a procurement decision.

Samantha: Expect pressure on Mistral, on Aleph Alpha, and on the broader European AI ecosystem to show they can compete at this frontier. Because if you can't build it, you either buy American or you regulate around it.

AI agents everywhere — and hackers are watching.

Bob: Staying with AI and security, because the two are really inseparable now. The so-called AI agent boom is genuinely here. Systems like OpenClaw reportedly have millions of users, and these aren't chatbots anymore. They book flights, move money, write code, click buttons on your behalf.

Samantha: And that last bit — click buttons on your behalf — is exactly where the cybersecurity community is losing sleep. Because once an agent can take actions, it becomes both a powerful tool and a very juicy target.

Bob: Security firms are saying the threat level is, quote, definitely much greater. Think about it: you give an agent access to your email, your calendar, maybe your bank. If an attacker hijacks that agent, or tricks it with a prompt injection hidden in a webpage, they don't just steal data — they take actions. They send money. They forward documents.

Samantha: And most users have no idea what permissions they've actually granted. Do you know how many tools your AI assistant has access to right now, Bob?

Bob: Honestly? I'd have to check. And that's the problem. The regulators are now being pushed to treat agentic AI as a distinct category, with its own rules around authentication, logging, and accountability when something goes wrong.

Samantha: Because right now, if your agent sends 50,000 euros to the wrong account, the legal picture is murky. Is it you? Is it the developer? Is it the platform?

Bob: Europe is actually reasonably well-placed on paper. The AI Act already has provisions for high-risk systems, and NIS2 covers cybersecurity obligations for essential services. But agentic AI is moving faster than the guidance documents.

Samantha: And that's the familiar European story, isn't it? Good frameworks, slow enforcement. Meanwhile, consumers in Berlin or Madrid are installing these agents today, not in two years when the guidance drops.

Bob: I'd expect to see stricter certification requirements coming, especially for agents that touch financial services or healthcare. And liability rules — that's the big one. If a European company deploys an American agent that goes rogue, who pays?

Samantha: My bet is the Commission accelerates work on an agent-specific annex to the AI Act. Because if they don't, national regulators will just start doing it themselves, and we'll end up with 27 different rulebooks.

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 moves to regulate AI avatars of real people.

Bob: Let's go to Beijing. China's cyberspace regulator has published draft rules aimed at the digital humans industry. These are AI avatars — many of them modelled on real people — that have become massively popular on Chinese social media, mostly for product promotion.

Samantha: And the scale is remarkable. You scroll through Chinese platforms and half the livestream hosts selling skincare or electronics are partly or wholly AI-generated. They work 24 hours a day, they never misspeak, and they're cheap.

Bob: The new rules focus on three things. Protecting children from harm, preventing avatars being created without consent, and a general, let's call it, social stability angle — which in Chinese regulatory language covers a lot of ground.

Samantha: The consent piece is the one I find most interesting globally. Because this is essentially deepfake regulation dressed up as industry rules. If you can't make an avatar of someone without their permission, that's a significant step.

Bob: And it's also been extending into less savoury corners — erotic chatbots built on real likenesses, for example. That's clearly part of what's driving this crackdown.

Samantha: It's one of those cases where China is actually moving faster on a specific AI harm than most Western democracies.

Bob: For Europe, this matters because the EU is still figuring out how to handle synthetic media under the AI Act. The Act requires disclosure of AI-generated content, but enforcement details are thin.

Samantha: And Europe has its own scandals driving this conversation. Politicians' voices being cloned, celebrity deepfakes on social platforms, romance scams using synthetic video. The harm is real and it's happening now.

Bob: What the Chinese approach shows is that you can regulate the creation step, not just the distribution step. That's a different philosophy. In Europe we tend to regulate platforms after the fact. Going upstream to the model providers and avatar creators is a more aggressive move.

Samantha: My guess is European policymakers will borrow selectively. The consent requirement, yes. The social stability framing, definitely not. But the basic idea that you need permission to digitally replicate a real person? I'd be surprised if that isn't codified in Europe within two years.

Jensen Huang is frustrated, and he's not hiding it.

Samantha: Moving to chips. Jensen Huang, Nvidia's CEO, went on a podcast and basically unloaded about US restrictions on selling AI chips to China.

Bob: And his argument is pragmatic rather than ideological. He's saying: look, China has plenty of computing power already. They're going to build their own AI ecosystem regardless of what Washington does. All the export controls achieve is pushing them to do it faster, and entirely outside the American tech sphere.

Samantha: His phrase was that he doesn't want the world to split into two tech systems. And if you're Nvidia, obviously, a unified global market is better for business. But he does have a point beyond the self-interest.

Bob: He does. Historically, when you cut a major economy off from your technology, they don't give up — they develop alternatives. Think about how Russia responded to sanctions, or how China built its own payment systems after being frozen out of Western ones.

Samantha: And Chinese chip champions like Huawei and SMIC have been making steady progress. Not at the absolute frontier yet, but closer every quarter.

Bob: Huang's view is that America wins by running faster, not by slowing China down. Keep innovating, keep the best chips in the American ecosystem by being genuinely the best, not by banning the competition.

Samantha: Whether the White House agrees is another matter entirely.

Bob: And this is where Europe gets squeezed. European AI companies, car makers, research labs — they all depend on Nvidia chips. If the global ecosystem splits, Europe has to pick a side, or try to build its own, which is very expensive and very slow.

Samantha: There are European efforts, of course. SiPearl, the Franco-European chip initiatives, the quantum push. But at the bleeding edge of AI training hardware, Europe is almost entirely dependent on imports.

Bob: So a fragmented tech world means higher prices, supply chain risk, and potentially being forced to comply with US export rules that restrict what European firms can do with their own products. That's a sovereignty issue dressed up as a trade issue.

Samantha: And I think European policymakers quietly agree with Huang, even if they won't say it publicly. A split world is worse for everyone who isn't Washington or Beijing.

AI meets ESG — and the numbers are staggering.

Samantha: Let's talk money. A new forecast projects that the market for AI in ESG and sustainability will hit 846 billion dollars by 2032. That's from about 8 billion today. It's a more than hundredfold increase in seven years.

Bob: Those numbers always deserve a pinch of salt, but the direction of travel is real. Companies are drowning in ESG reporting requirements — emissions data, supply chain audits, climate risk disclosures — and AI is genuinely useful for sorting through all of it.

Samantha: Generative AI apparently has over 41 percent market share in this space already. It's automating report generation, running scenario models, spotting anomalies in supplier data.

Bob: And it's driven, honestly, by regulation. This isn't companies suddenly getting green. This is companies getting compliance letters and realising they need software to keep up.

Samantha: A compound annual growth rate above 21 percent tells you this is one of the fastest-growing enterprise software categories right now.

Bob: The worry some people have, and it's a legitimate one, is that AI-generated ESG reports could become a kind of sophisticated greenwashing. If the machine is writing the report about the machine's own carbon footprint, we need very good auditing.

Samantha: Trust, but verify. Or in this case, automate, but audit.

Bob: Europe is the main customer here, frankly. The Corporate Sustainability Reporting Directive is forcing thousands of companies to produce detailed ESG reports, with external assurance. The compliance burden is massive, and AI is basically the only way to handle it at scale.

Samantha: Which also means there's an opportunity for European software firms. SAP, for instance, has been pushing heavily into sustainability tools. There's a whole generation of Nordic and Dutch startups building audit-ready AI for ESG.

Bob: And European regulators will have to think about how they trust AI-generated disclosures. If every company's sustainability report is being written by the same three models, you risk systemic blind spots.

Samantha: Still, on balance, this is one of those rare stories where regulation and innovation are genuinely pulling in the same direction.

DARPA wants to build a quantum computer that's a bit of everything.

Bob: And for our closer, something genuinely cool from the quantum world. DARPA has launched a program called HARQ — Heterogeneous Architectures for Quantum. The idea is simple, but the engineering is wild.

Samantha: So most quantum computers today pick one type of qubit and build around it. Superconducting qubits, trapped ions, photons, neutral atoms — each has strengths and weaknesses. HARQ wants to combine them in a single system.

Bob: Which sounds obvious when you say it out loud. Because that's exactly what classical computers do — you have CPUs for general work, GPUs for parallel tasks, specialised chips for AI, all talking to each other. Why wouldn't quantum eventually go the same way?

Samantha: The honest answer is because it's really, really hard. Keeping one type of qubit stable is difficult enough. Getting two different quantum systems to work together coherently, at near absolute zero, is a serious engineering challenge.

Bob: But if they pull it off, the applications are genuinely exciting. Better materials science, better drug discovery, better chemistry simulations. The kinds of problems where even small improvements change what's possible in the real world.

Samantha: And I love that they're borrowing a principle from classical computing. Sometimes the best ideas in a new field are old ideas from the old field, just applied differently.

Bob: It's a nice reminder that progress isn't always about invention. Sometimes it's about remixing what already works.

Samantha: And there's a direct European angle. One of the performer teams on HARQ is based in Switzerland. The UK has its own substantial quantum program. Germany, the Netherlands, France — all investing seriously.

Bob: Europe has real strengths in quantum. QuTech in Delft, the Paris-Saclay cluster, the Munich quantum valley. What HARQ signals is that the architecture race may be the next big frontier, not just the qubit race.

Samantha: So European teams will need to decide: do we join heterogeneous architectures, or do we double down on a single qubit type and try to win on depth? Probably a mix of both.

Bob: And this is a field where being second by two years can mean being irrelevant. So the pace matters.

Samantha: It's a fun story because it's genuinely forward-looking. No one's getting hacked, nothing's on fire. It's just clever people trying to build something remarkable.

Bob: And I'll take that as our Sunday note of optimism.

Samantha: Today we covered: Barclays warning on Anthropic's Mythos model, the cybersecurity risks of AI agents, China's new rules on digital humans, Jensen Huang's frustration with US chip restrictions, the booming AI market for ESG and sustainability, and DARPA's heterogeneous quantum computing program.

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.