Starlette BadHost Flaw Exposes Millions of AI Agents

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

Bob: And I'm Bob Russell. Today: Anthropic blows past OpenAI with a stunning new valuation, Apple hands the keys of Siri to Google's Gemini, Illinois passes a landmark AI safety law, a critical security flaw puts millions of AI agents at risk, CNN drags Perplexity into court over scraped articles, and a Blue Origin rocket goes up in flames during a ground test. Let's start with Anthropic.

Anthropic closes a sixty-five billion dollar round and hits a 965 billion dollar valuation, overtaking OpenAI.

Samantha: Anthropic has officially become the most valuable AI startup in the world. The company closed a Series H funding round of sixty-five billion dollars, pushing its post-money valuation to 965 billion. That puts it ahead of OpenAI, and frankly, ahead of nearly every public tech company on the planet.

Bob: And the timing is striking. This round lands just as Anthropic rolled out Claude Opus 4.8, the latest version of its flagship model. Early testers say it's noticeably more reliable, with fewer of those confident-sounding wrong answers we've all come to dread. Anthropic also introduced something it calls Dynamic Workflows, which lets Opus coordinate hundreds of smaller AI helpers to handle complex tasks in parallel.

Samantha: So instead of one big model doing everything, you have a swarm of specialised AI assistants working under one conductor. That's the direction the whole industry is moving in, but Anthropic is now extremely well funded to make it happen at scale.

Bob: The valuation itself tells you where investor money is flowing. We're talking about a private company worth nearly a trillion dollars. Three years ago that would have been unthinkable for a startup that hadn't shipped a consumer hit. The bet is that enterprise AI, the kind that quietly runs inside banks, law firms, and hospitals, will be where the real money lives.

Samantha: And there's a clear competitive signal here too. OpenAI was long seen as the default leader. Anthropic overtaking it in valuation doesn't mean it has overtaken it in users or revenue, but it does mean the market thinks the race is genuinely open.

Bob: For European listeners the picture is mixed. On one hand, this puts even more distance between American AI capital and European efforts like Mistral or Aleph Alpha. The combined valuation of Anthropic and OpenAI alone now dwarfs the entire European AI sector. On the other hand, Anthropic has been one of the more cooperative players when it comes to safety frameworks and the EU AI Act, so a bigger Anthropic could mean more leverage for European-style guardrails globally.

Samantha: The other angle is talent. With this much capital, expect another round of poaching from European universities and research labs. The pull factor towards San Francisco salaries just got stronger again.

Bob: And there's the infrastructure question. Sixty-five billion will fund enormous compute deals, which means more orders for Nvidia chips, more data centre construction, and more pressure on the power grid. Some of that build-out will land in Europe, particularly in the Nordics and Ireland, but the centre of gravity remains firmly on the other side of the Atlantic.

Apple is rebuilding Siri from scratch on top of Google's Gemini, with launch planned for iOS 27.

Bob: Story two, and it's a genuine surprise. Apple is reportedly rebuilding Siri from the ground up, internally code-named Campos, and the engine under the hood will not be Apple's own technology. It will be Google's Gemini.

Samantha: This is the biggest overhaul in Siri's history. The new version will ship as a standalone conversational AI app with iOS 27, with AI-powered web search built in. We're talking about more than a billion iPhone users worldwide getting access to a fundamentally different assistant.

Bob: What makes this so striking is the strategic admission. Apple has spent years insisting its own AI work was on track. Reaching out to Google, its biggest search rival, says the gap between Siri and what ChatGPT, Claude, and Gemini can do had become impossible to close internally on the timeline customers expect.

Samantha: It also deepens an already fascinating relationship. Google already pays Apple billions every year to remain the default search engine on the iPhone. Now Google's AI will be powering the assistant itself. That's a remarkable amount of Apple's user experience handed to one partner.

Bob: For European users this raises immediate regulatory questions. The Digital Markets Act treats both Apple and Google as gatekeepers, and a deal this deep between two of them will almost certainly trigger scrutiny from the European Commission. Data flows, user choice, and competition concerns all come into play.

Samantha: There's also a privacy dimension. Apple has built much of its brand on processing things on the device. Plugging into Gemini for serious conversational tasks means at least some queries will leave the iPhone and run on Google infrastructure. Apple will need to explain very clearly to European users what stays local and what doesn't.

Bob: The bottom line is that European iPhone owners are about to get a far more capable assistant, possibly later than the rest of the world if regulators ask hard questions first. And it forces every other handset maker, including Samsung in Europe, to rethink what their assistant offering looks like next year.

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.

Illinois passes a landmark AI safety law forcing frontier labs to publish risk plans and submit to annual audits.

Bob: Story three takes us to American state-level regulation, which is suddenly moving faster than Washington. Illinois has passed a bill known as SB three fifteen, and it sets what could become the new American standard for AI oversight. Governor JB Pritzker has signalled he will sign it.

Samantha: The law targets what it calls frontier AI companies, the biggest players like OpenAI, Anthropic, and a handful of others. They will be required to draw up, publish, and annually update detailed plans for how they handle serious or catastrophic risks from their models.

Bob: And here's the part that really matters: the law mandates annual independent safety audits. That's a first in American AI legislation. Until now, safety claims from the big labs have largely been self-reported. Illinois is saying that's no longer enough.

Samantha: It's a fascinating echo of what Europe is already doing under the AI Act. Brussels has been pushing for exactly this kind of structured risk assessment and external review for high-risk and general-purpose AI systems. Illinois has effectively imported the European playbook into the American Midwest.

Bob: For European policymakers that's a quiet win. One of the criticisms of the AI Act has been that it would isolate European companies if no one else adopted similar rules. If a major American state is now requiring comparable transparency, the compliance overhead for global labs becomes less of a competitive disadvantage for European firms.

Samantha: It also raises pressure on other US states. California has been working on its own version, and a patchwork of state-level rules could push the big AI labs to ask Washington for federal legislation just to simplify things.

Bob: The catch is enforcement. Publishing a risk plan is one thing, having auditors with the expertise to actually evaluate frontier models is another. Europe is wrestling with the same shortage of qualified reviewers. Illinois will need to build that capacity, and quickly.

A critical flaw called BadHost in the Starlette framework leaves millions of AI agents exposed to attackers.

Samantha: Story four is one of those underreported stories that really should be louder. Security researchers have uncovered a critical vulnerability in Starlette, the Python framework that sits underneath some of the most widely used AI tools in the world.

Bob: We're talking about FastAPI, vLLM, and LiteLLM, all built on top of Starlette. The flaw, nicknamed BadHost, lets an attacker bypass authentication by injecting a single character into an HTTP header. One character. That's all it takes to walk past the front door of vulnerable systems.

Samantha: A patch was released on May twenty-first, version 1.0.1, but the problem is that vulnerable versions are still running everywhere. Many AI deployments are built once and rarely updated, which means millions of AI agents are sitting exposed right now.

Bob: This matters enormously because of the direction the industry is heading. Just a few minutes ago we talked about Anthropic launching tools that coordinate hundreds of AI helpers. A lot of those helpers run on exactly the frameworks affected by this flaw. An attacker getting into one agent could potentially hijack a whole chain of them.

Samantha: For European companies, the practical question is uncomfortable. Most enterprises adopting AI right now don't actually know which framework underlies the tools their vendors sold them. So the call to action is straightforward: ask your IT teams whether Starlette is in your stack, and whether it has been patched to 1.0.1.

Bob: There's also a policy layer. The EU's NIS2 directive and the upcoming Cyber Resilience Act both put obligations on companies to keep software dependencies up to date. A flaw this severe, sitting unpatched in production, could increasingly become a regulatory problem, not just a technical one.

Samantha: It's a reminder that the AI boom rides on a stack of open source plumbing maintained by relatively small teams. One header parsing bug, and the whole thing wobbles.

CNN sues Perplexity for allegedly copying paywalled articles word for word and serving them to users.

Bob: Now to something a bit more practical you can keep an eye on today. CNN has filed a lawsuit against AI search startup Perplexity, alleging that the company scraped paywalled articles and served verbatim copies to its users.

Samantha: What makes this case different from earlier AI copyright disputes is the specificity. CNN isn't arguing the vague fair-use territory that has dominated lawsuits against OpenAI or Anthropic. It claims Perplexity literally copied and redistributed full passages of its journalism.

Bob: For listeners, here's the practical angle. If you use Perplexity as your daily search tool, this lawsuit could change what you see. Perplexity has built its reputation on giving direct, sourced answers, often with substantial quotes from news sites. If CNN wins, expect that experience to change quite quickly.

Samantha: European publishers will be watching this closely. Groups representing newspapers in France, Germany, and the Netherlands have been pushing for licensing deals with AI companies, and a CNN victory in American courts would strengthen their hand enormously in negotiations.

Bob: There's also the EU copyright directive, which already gives publishers more rights over how their content is reused online. A clear American precedent that verbatim scraping crosses the line would give European regulators a lot of confidence to enforce more aggressively at home.

Samantha: The practical takeaway for listeners using AI search tools today: pay attention to whether the answers you get are summarised or quoted. The economics of AI search depend heavily on what courts decide here.

Bob: And if you're a journalist, blogger, or run any kind of content site in Europe, this is a good moment to check your terms of service and your robots.txt file. Make explicit what AI crawlers are and aren't allowed to do. It costs you nothing today, and it strengthens your position later.

Blue Origin's New Glenn rocket explodes during a ground test at Cape Canaveral.

Samantha: And to close out, something concrete you can actually follow today and a clear consumer-level story. Jeff Bezos's space company Blue Origin had a bad night at Cape Canaveral. Its heavy-lift New Glenn rocket exploded during a prelaunch hotfire test.

Bob: The blast was visible from miles away. Thankfully all personnel are accounted for and no injuries have been reported. But the damage to the launch site is significant, and the investigation into what went wrong will take time. Future launches are almost certainly delayed.

Samantha: For the average space enthusiast, there's something concrete to do today. Blue Origin streams its test footage and updates publicly. You can follow the investigation live through their channels, and amateur trackers on YouTube and on sites like NextSpaceflight will be reconstructing what happened minute by minute over the weekend.

Bob: For European listeners interested in the bigger picture, the implications matter. New Glenn was supposed to be a serious competitor to SpaceX's Falcon Heavy. A delay strengthens SpaceX's near-monopoly on heavy lift, which is uncomfortable for Europe's own Ariane 6, which already struggles to compete on price.

Samantha: It also affects customers, including Amazon's Kuiper internet satellites, which were counting on New Glenn for a chunk of their launches. Expect some of that demand to shift back to SpaceX, or to European providers if they can scale up fast enough.

Bob: So today: a smart find for space fans is to bookmark the NASA Spaceflight forums and the Blue Origin press page. You'll get raw, unfiltered updates as the investigation unfolds, well before mainstream coverage catches up.

Samantha: Today we covered: Anthropic overtaking OpenAI on valuation, Apple rebuilding Siri on Gemini, Illinois passing a landmark AI safety law, the BadHost flaw exposing millions of AI agents, CNN suing Perplexity over scraped articles, and Blue Origin's New Glenn explosion at Cape Canaveral.

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.