White House AI Safety Tests: Labs, Cyber Rules & TSMC

Samantha: Welcome to The State of Tech, The European Edition, Tuesday August fourth, 2026. I'm Samantha Lawrence.

Bob: And I'm Bob Russell. Today: the White House sits down with the biggest AI labs to hammer out safety tests, CrowdStrike warns that AI has moved to the centre of every serious cyberattack, TSMC pushes for a fivefold jump in its most advanced chip production, Stanford calls for urgent rules on so-called world models, and to close out, two things you can actually use today: the free open-source translator LibreLingo, and the Danish encrypted note app Turtl. Let's start with the White House.

The Trump administration brings OpenAI, Anthropic, Google and Meta together to draft voluntary AI cybersecurity safety tests.

Samantha: The Trump administration is hosting the biggest names in artificial intelligence at the White House today. OpenAI, Anthropic, Google and Meta are all at the table, and the subject is a new framework for voluntary safety testing of advanced AI models. The specific focus, according to officials, is measuring how good these models are at hacking. That flows from a June executive order on AI cybersecurity, and it lands at a moment when concerns in Washington have shifted from theoretical harms to very concrete ones.

Bob: And the trigger is not abstract. In recent weeks both OpenAI and Anthropic have disclosed that their own AI tools breached other companies' systems during authorised cybersecurity tests. In other words, the labs themselves are showing that their models can already break into things. Lawmakers on both sides in the US are worried that the same capability, in the wrong hands, becomes a mass-scale cyberattack tool. The benchmarks the White House wants to build have not been published yet, so it is still framework talk rather than hard rules.

Samantha: The word to underline here is voluntary. The US is not going for binding legislation, it is going for a shared testing standard that the labs sign up to. That is a very different model from Europe, where the AI Act sets legal obligations, and it raises the question of whether a voluntary regime can keep pace when the capabilities are moving this fast.

Bob: For European regulators, this matters in two ways. Any benchmark the US labs adopt effectively becomes the global baseline, because these are the same companies shipping into Europe. And on cybersecurity, the infrastructure is shared, an AI-driven attack does not stop at the Atlantic. So Brussels will watch the wording carefully, especially where voluntary US commitments start to overlap with binding European obligations under the AI Act and existing cyber rules. The bottom line: the frame being drafted in Washington today will shape the conversation European supervisors have tomorrow.

CrowdStrike's new threat report says AI is now weapon, target and force multiplier for cyber attackers.

Samantha: CrowdStrike released its 2026 Threat Hunting Report yesterday, and the headline finding is blunt: artificial intelligence is now embedded across modern adversary operations. Not as a novelty, not as a future risk, but as a working part of how attacks are planned, executed and scaled. The report tracks how China-linked groups are exploiting critical vulnerabilities within twenty-four hours of a public proof-of-concept appearing online. That is faster than most companies can even schedule a patch meeting.

Bob: And it is not only speed. North Korea-linked groups have reportedly poisoned one hundred and thirty-one trusted AI framework packages. That means the tools developers pull in to build their own AI systems can already carry hidden malicious code. It flips the picture. Until recently, AI was mostly discussed as something companies defend. This report says AI infrastructure itself is now the battleground, from the software supply chain to enterprise language models to the AI workloads running in the cloud.

Samantha: The practical worry for a European IT team is compound risk. You adopt AI to be more productive, but every new AI component is also a new door into the network. Enterprise chatbots get abused with prompt tricks, model weights sit in cloud buckets that attract targeted access attempts, and a poisoned package can quietly reach thousands of downstream customers before anyone notices. CrowdStrike's argument is that the defenders now need AI-driven detection just to keep pace.

Bob: For European businesses and public bodies, that pushes AI supply chain security up the priority list, alongside classic questions like patching and identity. Expect this report to be quoted a lot when procurement teams start asking harder questions about where their AI components actually come from, and what has been done to verify them.

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.

TSMC targets a fivefold jump in two-nanometre chip production by the end of 2026 as AI demand explodes.

Bob: TSMC, the Taiwanese company that manufactures most of the world's most advanced chips, is ramping up its two-nanometre production faster than almost anyone expected. The target is one hundred thousand of these advanced wafers per month by the end of 2026. Right now this generation of chip accounts for only around three percent of the company's revenue, so we are talking about a roughly fivefold increase in output in a very short window. The reason is simple: AI customers are lining up.

Samantha: NVIDIA and AMD are the names most often mentioned as securing capacity, because their next generation of AI accelerators depends on this node. And this is where the chip story becomes an energy story, a geopolitics story, and a supply story all at once. Every big Western cloud provider planning new AI data centres is effectively planning around what TSMC can deliver on this exact line.

Bob: For Europe, there are two angles worth pulling apart. The upside is access: European carmakers, industrial groups, and AI startups all rely, directly or indirectly, on chips coming off these lines. More capacity means fewer bottlenecks and, over time, more competitive pricing on high-end silicon. The downside is concentration. The most strategically important process in the industry sits on one island, and Europe's own chip ambitions, including the Chips Act projects still under construction, are years away from touching this level of sophistication.

Samantha: So the message is mixed. In the short term Europe benefits from TSMC pushing hard. In the medium term, this ramp is a reminder of just how far European sovereignty ambitions still have to run before the continent can meaningfully produce its own leading-edge AI chips.

Stanford researchers urge urgent rules for world models, the AI systems that map the physical environment around us.

Bob: Researchers at Stanford's Institute for Human-Centered AI have published a policy brief warning that a new class of AI, so-called world models, needs urgent regulatory attention. Where a language model predicts the next word, a world model tries to build an internal picture of a real physical environment and predict what will happen next in it. Think self-driving cars anticipating traffic, robots planning movements, or crisis response systems modelling how a fire or a flood will spread.

Samantha: The Stanford team's argument is straightforward: if governing large language models has already stretched regulators, world models will be much harder. These systems are moving from research labs into commercial products right now, and they touch domains where mistakes have physical consequences, not just misleading text. The brief flags implications for privacy, because these models hoover up sensor data from cameras and lidar, and for national security, because the same technology underpins autonomous vehicles and, potentially, autonomous weapons.

Bob: What makes this different from the usual AI regulation debate is the closing window. The researchers argue that the moment to set the rules is before these systems are deeply embedded in cars, factories and public infrastructure, not after. Once they are woven in, retrofitting oversight becomes politically and technically much harder.

Samantha: For Europe, the practical question is whether the AI Act, which was largely drafted around generative and predictive systems, is enough. Legal scholars are already asking whether world models slot cleanly into the existing high-risk categories, or whether a specific track is needed. Expect this brief to feed directly into that debate in Brussels over the coming months, especially as the first commercial deployments in mobility and emergency response arrive on European roads.

LibreLingo is the free open-source language app that keeps your learning private and works offline.

Bob: To close out, two things worth trying today. First up is LibreLingo, a free open-source language learning app that has been quietly building a following as a genuine alternative to the big commercial platforms. It offers structured courses in a growing number of European languages, uses spaced repetition to lock vocabulary in, and it does not need an account or a subscription to get started.

Samantha: The appeal is not just the price. LibreLingo does not chase you with streak notifications or push you toward paid tiers, and because it is open source, the code and the course content are transparent. Users have contributed courses in Spanish, French, German and several smaller European languages, and the community is actively building more. For anyone who has felt slightly hostage to a single big language app, this is a plainer alternative that focuses on the actual lessons.

Bob: Practically, it runs in a browser, and there are mobile builds available depending on your device. Sessions are short, usually five to ten minutes, which fits nicely into a commute or a coffee break. Because everything is on your side, your learning history is not being fed into an ad model, which matters for anyone teaching kids a second language and not thrilled about their progress data ending up in a marketing pipeline.

Samantha: It is not going to replace an intensive course or a good teacher, but as a free daily habit it holds up well. If you have been thinking about brushing up your French before a summer trip, or finally getting past tourist German, LibreLingo is worth an afternoon of your time. And because it is open source, if a course you want does not exist yet, you or a colleague can help build it.

Turtl is the free Danish encrypted note app that keeps your ideas genuinely private across every device.

Bob: The second thing to try today is Turtl, a free note-taking and bookmarking app built by a small Danish team. Turtl encrypts everything you write on your device before it ever leaves it, so notes, links, files and photos sync between your laptop and your phone without the servers being able to read a word of them. The company explicitly cannot see your content, and the source code is open for anyone to inspect.

Samantha: What makes Turtl different from just another notes app is the structure. It uses boards rather than folders, so you can sort research, personal thoughts, and work notes into separate spaces, and you can share individual boards with other Turtl users while keeping the rest private. That is useful for anyone collaborating on a sensitive project without wanting to hand the entire notebook to a big cloud provider.

Bob: It works on Windows, Mac, Linux and Android, and there is a browser version too. Setup is minimal, and because it is designed by a European team with privacy as the starting point rather than an afterthought, it fits well alongside other privacy-first tools. Reviewers highlight that the interface is calm and uncluttered, which for a notes app matters more than any fancy AI feature.

Samantha: A small caveat: because everything is encrypted client-side, if you lose your password there is no recovery. So write it down somewhere sensible. But for anyone who has been looking for a European alternative to the mainstream note apps, and who wants their thinking to stay their own, Turtl is a solid pick to install today.

Samantha: Today we covered: the White House pulling the big AI labs together on cybersecurity testing, CrowdStrike's warning that AI is now central to modern cyberattacks, TSMC's aggressive push into two-nanometre chip production, Stanford's call for urgent rules on world models, and two things to try yourself: LibreLingo and Turtl.

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.