Armadin raises $255M for AI attackers & Deepslate voice AI

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

Bob: And I'm Bob Russell. Today: the US competition watchdog opens a sweeping probe into OpenAI and Anthropic over their safety claims; Planet Labs puts the first demo of Google's orbital AI chips into space; OpenAI fires three safety researchers after an internal leak; the Eclipse Foundation launches a Sovereign AI push for Europe; and to close out, two things worth trying today, the open-source read-later app Wallabag and the free European mapping service Mapy.cz. Let's get started.

A Mandiant Founder Raises $255 Million to Build Robot Hackers

Samantha: Okay, let's talk about a company that pays serious money to attack other companies. Armadin, led by Kevin Mandia, the guy who founded Mandiant, just raised two hundred fifty-five and a half million dollars in a Series B.

Bob: At a valuation north of two and a half billion. And it only came out of stealth seven months ago. That's a very fast climb.

Samantha: It is. And the product is basically autonomous AI attackers. Software that behaves like a determined intruder, probes your network, finds the weak spots, and then tells you what it found.

Bob: Which is red teaming, but without the weeks of scheduling and the five-figure daily rate for a human crew. You can run it constantly instead of once a year before the auditor shows up.

Samantha: That's the pitch, yeah. Defence has always been a point-in-time exercise, and attackers aren't working point in time. They're working every day.

Bob: Here's what nags me, though. We've spent the last few weeks hearing about how AI agents can be hijacked with a single email, and how AI-found vulnerabilities are twice as likely to allow remote code execution. So you're building an agent that's good at breaking in. How do you keep that thing on a leash?

Samantha: Good question, and I'd guess that's where a lot of that money goes. Scope controls, kill switches, proof that it only touches systems the customer authorised. If an autonomous attacker wanders outside its brief, you've got a real incident, not a test.

Bob: And the dual-use problem is obvious. The same capability that finds your holes for you is something a criminal group would love to own. The founder's track record helps with trust, but it doesn't change the underlying tech.

Samantha: True. Though I'd say the criminals are going to build this stuff anyway. The question is whether defenders get the same tools. Mandia's whole career is incident response, so he's seen how fast the other side moves.

Bob: That's the argument. I'd still want to see independent evidence that it finds things human testers miss, rather than just generating a very long report of low-grade findings.

Samantha: Fair. Noise is the killer in security tooling. If the output is a thousand alerts and ten matter, nobody wins.

Bob: But a quarter of a billion dollars says investors think the demand is there. Companies are drowning in attack surface, and they can't hire enough people to cover it.

Samantha: So we'll watch whether this turns into a standard line item in security budgets, or whether it's one more expensive scanner with a clever name.

Berlin's Deepslate Wants Its Own Voice AI, and Europe Is Cheering

Bob: Smaller story, but one I like. A Berlin lab called Deepslate just raised seven point seven million euros to build its own speech-to-speech model.

Samantha: Backed by 42CAP, Alstin Capital, SIVentures and some angels. A seed round, so we're talking early days.

Bob: Early, yes, but the technical choice is the interesting bit. Most voice assistants today work as a relay race. Your speech gets turned into text, a language model reads the text and writes a reply, then another system turns that reply back into speech.

Samantha: Which is why so many of them feel laggy. Every handoff costs you time, and you lose the tone, the hesitation, the sarcasm. All the stuff that makes speech speech.

Bob: Exactly that. A true speech-to-speech model skips the text middle step. Audio goes in, audio comes out, so it can respond faster and pick up on how you said something, not just what you said.

Samantha: In theory. In practice, building one is hard and expensive, and seven point seven million doesn't go far against the labs spending billions. I'd be curious what they can realistically train with that.

Bob: That's the honest worry. But a seed round buys you a prototype and a few customers, and the bet is that a focused team can do one thing well. Plus, they're in a part of the market where Europe has a real argument.

Samantha: Go on.

Bob: Languages. European companies deploying phone agents need German with a Swabian accent, Dutch, Polish, Portuguese. The big American models are strongest in English and then thin out. A European lab can treat that as the main job rather than an afterthought.

Samantha: And there's the data question. Call centres handle health details, banking details. If a voice agent is listening to all of it, plenty of companies would rather the audio never leaves Europe.

Bob: That's a selling point you can't easily copy from California. Though it only works if the quality is competitive. Nobody buys sovereignty if the agent mishears their address.

Samantha: Right, the customer in the queue doesn't care where the servers sit. They care whether it understood them the first time.

Bob: So Deepslate has to win on quality first, and then use the European angle as the tiebreaker. Worth watching.

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.

OpenAI fires three safety researchers over an alleged leak to an outside AI safety group.

Bob: OpenAI has dismissed three researchers from its safety team after an internal investigation concluded they passed confidential information to a third-party AI safety organisation. According to the company, the employees bypassed established procedures and mishandled sensitive material, which OpenAI calls a breach of trust. The company has not named the external group, and the fired researchers have not publicly responded in detail. What is clear is that this sits inside a longer pattern of tension between OpenAI's commercial trajectory and parts of its own safety staff.

Samantha: This is the second OpenAI safety story on the show today, and that is not a coincidence. The company is under formal investigation in Washington over its safety claims, and at the same time it is sanctioning people internally for allegedly going outside the official channels to raise concerns. From OpenAI's side, that is a straightforward security and governance matter. From the perspective of outside researchers and former staff, it reinforces a worry that internal dissent is being managed rather than welcomed.

Bob: There is a real policy question underneath this. If safety researchers at a frontier lab believe something is going wrong, where exactly are they supposed to go? Right now there is no formal protected channel to an independent regulator, the way there is for financial whistleblowers in some jurisdictions. The European AI Act contemplates reporting obligations for serious incidents, but not yet a clear safe harbour for individual employees who want to flag concerns externally.

Samantha: For European regulators this incident is going to feed directly into debates about corporate accountability and transparency under the AI Act. The bottom line is simple. The more decisions about advanced AI stay behind non-disclosure agreements, the harder it is for any outside body to independently verify safety claims, which is exactly what today's other open letter from more than one hundred twenty global leaders is calling for.

Europe's Eclipse Foundation launches a Sovereign AI Initiative for open, governed AI innovation.

Bob: The Eclipse Foundation has launched what it calls a Sovereign AI Initiative, aimed at fostering open and governed European innovation in artificial intelligence. The foundation is best known for stewarding major open-source projects used across industry, and it is now trying to apply that same model to AI. The focus is on tools, infrastructure and governance frameworks that align with European values and regulations, with transparency built into how models and data pipelines are developed.

Samantha: The word "sovereign" is doing a lot of work here. In practice it means reducing dependence on a handful of US and Chinese providers for the core building blocks of AI. Not by banning anyone, but by making sure there are credible European-governed alternatives for things like model training frameworks, evaluation tools and data-sharing standards. For a car maker in Germany or a hospital group in France, that matters, because today a lot of their AI stack rests on infrastructure they do not control.

Bob: The challenge is familiar. Europe has launched a lot of sovereignty initiatives in cloud, chips and now AI, and the gap with the hyperscalers remains large. What this initiative has going for it is a concrete institutional home and a track record in open source. Governance and openness are areas where Europe genuinely leads, and tying them explicitly to the AI Act gives companies a reason to engage beyond just goodwill.

Samantha: For listeners working inside European companies, the practical angle is that this creates another route to influence the standards their suppliers will have to meet. Industry working groups, reference implementations, shared evaluation benchmarks, that is where the real power sits. If the Eclipse effort attracts serious contributors from banks, telecoms and public sector buyers, it has a shot at shaping what trustworthy AI actually looks like in day-to-day European procurement.

Two things worth trying today, starting with the open read-later app Wallabag.

Bob: To close out, two things you can actually use today. First up is Wallabag, a free open-source read-later app from France. The idea is simple. You install a browser button, and whenever you find an article you want to read properly later, you click once and Wallabag strips out the ads, pop-ups and sidebars, and saves a clean version to your personal library. You can then read it on your phone, tablet or laptop, online or offline, including on an e-reader.

Samantha: What makes Wallabag stand out from the big American read-later services is where your data lives. You can either pay a small annual fee for the hosted European version, which runs on servers in France, or if you are technically minded you can install the software on your own machine or a cheap home server, and nothing leaves your house. The app is open source, so independent developers can inspect what it actually does with your reading list. Reviewers praise the clean reading mode and the fact that it handles paywalled newsletters and long reports without choking.

Bob: It also plays nicely with the rest of your reading life. Wallabag can import articles from browser bookmarks, from RSS feeds, and from other read-later tools if you are migrating away. There are mobile apps for Android and iOS, and community-built extensions for every major browser. For listeners who are drowning in open tabs and promise themselves they will read that long piece on EU chip policy "soon", this is a way to actually make that happen without feeling watched.

Samantha: A practical note on the limits. The hosted version is a paid subscription, modest but not zero, and the self-hosted option does require a bit of setup if you have never run your own server. For readers who just want the simplest possible button-and-cloud experience, there are easier options. But if you care about keeping your reading list out of a US advertising ecosystem, and you want something that will still work in five years regardless of what any one company decides, Wallabag is a solid, European-grounded pick for a Friday afternoon install.

Mapy.cz is a free European mapping and navigation service from the Czech Republic, with excellent hiking and cycling maps.

Bob: Pick number two is Mapy.cz, a free mapping and navigation service from the Czech company Seznam. You open it in any browser, or install the free app on your phone, type in where you want to go, and get routes for driving, cycling, walking and public transport. It looks and feels familiar if you have used Google Maps, but its real strength is outdoor mapping. The tourist, winter and aerial map layers are among the most detailed you will find anywhere in Europe.

Samantha: Under the hood, Mapy.cz builds on OpenStreetMap, the collaboratively maintained map of the world that many European public services already rely on. That means the geographic data is open, widely used and genuinely global, not locked inside a single company's walled garden. And Seznam is a European company, which matters for many listeners when they think about who runs the service and where it is based.

Bob: Practical use. You can use it straight away in your browser on a laptop or phone, and the mobile app is free as well. The app lets you download maps for offline use, which is handy when you are walking in the mountains or travelling abroad and want to save mobile data. For planning a weekend trip, checking a hiking route before you leave, or finding a cafe in a city you do not know, it is more than enough. Users praise the clear marked trails and the cycling routes, and coverage is especially strong in Central Europe.

Samantha: There are limits to be honest about. Some of the richer features of the big incumbents, like detailed indoor maps or real-time traffic on every road, are not at the same level everywhere, and the depth of detail is greatest in the Czech Republic and its neighbours. Also, you may want to read the privacy settings in the app and decide what you are comfortable sharing. But as a free, outdoor-friendly alternative for everyday European navigation, it is well worth bookmarking today.

Samantha: Today we covered: A Mandiant Founder Raises $255 Million, Berlin's Deepslate Wants Its Own Voice AI, OpenAI fires three safety researchers over, Europe's Eclipse Foundation launches a Sovereign, Two things worth trying today, en Mapy.cz is a free European mapping.

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.