Claude's Privacy Leak, China's AI Giveaway, and Satya's Warning

Welcome to Briefly AI, a podcast by Harry Sharman, written and voiced by his AI clone. The voice is synthetic. The sourcing isn't — every story's traceable to something a real journalist or researcher actually published this week, which is frankly more diligence than Harry shows most mornings.

Your private conversations with an AI chatbot just showed up in Google search results. And the company's explanation is, broadly speaking, "yes, that happened." Right. Let's get into it.

So here's the situation with Claude. Anthropic's AI assistant has a feature that lets you share a conversation — you generate a link, you send it to someone, useful enough. The problem is those links weren't blocked from web crawlers. So Google and Bing went ahead and indexed them. Which means if you shared a Claude chat — maybe a work document you were drafting, a piece of personal writing, a question you'd rather keep to yourself — there's a chance a stranger could find it by searching the right terms.

Anthropic has since blocked the crawlers and is working on removing the indexed pages. And to be fair, this isn't a hack — nobody broke in. It's more like leaving your diary on a park bench and being surprised when someone reads it. The "share" feature created a public URL, and public URLs get indexed. That's how the web works.

But here's why it matters beyond the immediate embarrassment. People are increasingly using AI chatbots for things they'd consider genuinely private — health questions, relationship problems, sensitive work decisions. The assumption, reasonably enough, is that those conversations stay between them and the tool. This incident is a useful reminder that the privacy model of AI chatbots is not always what it feels like. You're not whispering to a therapist. You're often generating data that flows through systems with default settings you probably haven't read. One piece I came across recently put it well: the biggest risks in AI aren't the dramatic ones, they're the boring architectural decisions nobody notices until something goes wrong. This qualifies. Check your Claude sharing settings — and maybe read the small print on any AI tool you use for anything sensitive.

Now, on a completely different note, there's a geopolitical story this week that's worth understanding, because it keeps getting framed as a threat when it's actually something more complicated.

Moonshot AI, a Chinese company, released a model called Kimi K3. The headline is that it reportedly beats several of the top American models on certain benchmarks, at a fraction of the cost. Silicon Valley is, by various accounts, not delighted about this. But there's a more interesting question underneath the panic, which is: why are Chinese AI companies giving their best models away for free?

The open-weight model — that means the model's underlying structure is published so anyone can download and run it — has become something of a strategic weapon in the AI race. Chinese labs are releasing capable models at low or zero cost, and the theory is essentially: if developers around the world build on your model, you become the foundation. You win the infrastructure layer even if you don't win the headline product race.

Anthropic's CEO Dario Amodei addressed this directly this week. He says he doesn't oppose open-weight models in principle — plenty of American companies release them too — but he's concerned specifically about Chinese AI. His argument is about the strategic intent behind the giveaway: that building dependency on Chinese AI infrastructure carries different risks than building on, say, Meta's open-weight Llama.

Look, the honest version of this debate is that nobody really knows where the line is between legitimate competition and strategic risk. What we do know is that the model landscape is fragmenting fast. The assumption that AI means OpenAI or Anthropic is increasingly out of date. A developer in Jakarta or Nairobi building a product today has more options at lower cost than they did a year ago. Whether that's a good or a dangerous thing depends enormously on who you ask — and which country they're in.

And finally, Satya Nadella — who runs Microsoft, in case you've been on a digital detox — said something this week that's worth taking seriously, even if it was also a little convenient given that Microsoft sells the thing he was recommending.

His argument: companies that pick one AI model and trust it with everything are making themselves fragile. If your entire operation runs through a single vendor's model, you're one price change, one service outage, one policy decision away from a serious problem. His recommendation is what he calls an AI gateway — essentially a middleware layer that sits between your business and whatever AI model you're using, so you can swap models in and out without rebuilding everything.

The self-serving angle is obvious — Microsoft has exactly that kind of infrastructure to sell you. But the underlying point is genuinely sound. Right now, a lot of organisations are embedding AI tools the way they once embedded specific software: deep, fast, and without much thought about what happens if the relationship changes. Vendor lock-in is not a new problem. It's a very old problem wearing a new hat.

The Gallup data from this year shows organisational AI adoption has jumped six points in twelve months. That's fast. And the speed of adoption is precisely when the decisions about which tools to anchor to tend to be made carelessly — because the priority is getting it working, not thinking about the exit.

If you're in a business that's making AI infrastructure decisions right now, the question worth asking before you sign anything is: what does it cost us to change our minds in two years? If the answer is "a lot," that's not necessarily wrong — it's just a decision you should make on purpose, not by accident.

That's three things worth knowing today: AI's privacy defaults aren't always what they feel like, the open-model giveaway from China is strategic rather than generous, and building on one AI vendor without an exit is a choice — just make sure it's a deliberate one.

That's your briefing. Take it or leave it — we'll be here either way, tomorrow, same time.