A security researcher hid two million fake secrets inside AI prompts and then watched where they turned up. Some came back from hackers. Some came back from the very middlemen selling the AI access. If you paste customer chats into a cheap AI tool, this is the story to read before your next paste.
Under it today: Walmart is building a front door for other people's shopping agents, Signal has started blocking phone assistants from reading chats, and ChatGPT is redesigning its phone app around an always-on agent. Sorted, as usual, by how close each one gets to your shop.
Two Million Bait Prompts Show Where Your AI Data Really Goes
HD Moore, chief research officer at the security firm Dragos, ran the test from July to November. He sent close to two million prompts through popular AI services, each one carrying a unique fake password, key or link. Then he waited to see which ones got used.
Three things came out of it, according to Ars Technica and Moore's own summary. Three hosting providers had been breached: attackers lifted the bait cloud keys and tried them in tens of thousands of calls, sometimes within hours. Some resellers were reading prompts themselves, by offering low prices, switching on logging and waiting. And several providers ignored privacy settings such as "zero data retention", the flag that is supposed to mean nothing is kept. OpenRouter, a large routing service, has reportedly pulled that label from some endpoints while it investigates.
Here's why I put this first. Most small sellers never buy AI straight from the big labs. They buy it through an app, which buys it through a router, which may buy it from a discount reseller. Your prompt gets opened at every stop. And what's in a seller's prompt? Customer names, phone numbers, addresses, order history, sometimes a login someone pasted in a hurry.
Nobody is saying every cheap tool is a thief. The finding is narrower and more useful: the weak points are the middle layers, and a privacy label is only a promise. So ask your vendors three plain questions.
- Who processes my prompts? A direct contract with a first-tier provider, or a chain of resellers?
- What gets logged, and for how long? "We don't keep anything" should come with a document, not a shrug.
- What never goes in? Passwords, API keys and card details should never be in a prompt, whatever the tool.
It's the same principle I covered when banks made customers liable for their AI agents: the cost of a leak lands on the account owner. That's you.
Walmart Builds a Front Door for Other People's Shopping Agents
On its earnings call this week, Walmart said US use of its Sparky shopping assistant grew 129% in a year, and that baskets tied to Sparky were 40% larger. Good numbers. The more interesting part came next.
Walmart is opening an Agent Gateway, in beta for businesses and developers. It lets outside AI agents shop on a customer's behalf without seeing personal details or purchase history. Next month it plans "Agent Profiles", where shoppers set spending limits, track what an agent bought, and cut off access. Amazon, for contrast, has been trying to block outside agents from shopping on its site at all.
So the two biggest retailers in America now disagree on the main question of the next few years: do you let the robots in? Walmart's answer is yes, through a checked door, on its terms.
For a small seller the lesson isn't "build a gateway". It's that agents buy from stores they can read. Clear product names, real prices, current stock, a straight answer on delivery. A human who gets "DM for price" might write to you. An agent moves on. If your catalogue lives in your head and your Stories, start with what actually works in AI customer service for online stores and fix the basics first.
Signal Shuts the Blinds on Phone Assistants
Signal announced on November 20 that it will block screenshots by default in its mobile apps. Not to stop your friends. To stop the AI assistant built into the phone from reading the screen.
On iOS 27 there is a new "Block Siri AI" setting, on by default. On Android, Signal's screen security now switches on by itself when it detects an assistant trying to read the display. Signal admits the protection is incomplete, says it can't reliably do the same on Mac, and is asking Apple, Google and Microsoft for a simple way for apps to opt out of AI access. The move follows Reuters tests showing Meta's Muse agent could pull Signal message content from notifications and screenshots.
Few Iranian sellers run their shop on Signal, so why does it matter? Because the same assistants sit on the phone where you answer Instagram and WhatsApp. If an assistant can read one chat app, it can read the others, and neither of those has pulled the blinds yet. A customer's address in a DM is now visible to more software than you think. Worth a look at what the assistant on your work phone is allowed to see.
ChatGPT Redesigns Its Phone App Around an Agent
OpenAI began testing a new ChatGPT mobile layout on November 19. The headline change: Dot, its always-on agent, gets its own tab, next to a bottom bar where you can pin tools like Projects and Images. For now it's a test for Pro users in eligible regions, with a switch in settings to turn it off. OpenAI gave no date for a wider release.
A tab is a small thing. It's also how a company tells you what it thinks the product is. A year ago ChatGPT was a box you typed questions into. Now the box shares the screen with something that works while you aren't looking.
My take for sellers: your customers are being trained, app by app, to hand a job to an assistant and expect it done. That expectation doesn't stay inside ChatGPT. It follows them into your DMs, where "I'll get back to you tomorrow" already sounds slow. If you want to see how wide that gap is, I'd read AI versus human reply speed: 9 seconds against 60 minutes.
Quick takes
- Claude can now drive desktop apps for teams. Anthropic's release notes say computer use is in beta on Team and Enterprise plans, rolling out gradually. It is off by default and an owner has to switch it on. Sensible, given it works with no sandbox between the AI and your apps.
- Target has an assistant too. "Ask Target" is now in beta, per CIO Dive. Every big retailer wants its own agent in front of the customer before somebody else's gets there.
- The UN's AI safety summit is tomorrow. Barbados and the UN host it in Bridgetown on November 23. Don't expect rules: the UN says it won't create global AI regulation. A separate, non-binding call to control frontier AI models now has 32 states and the European Commission behind it, per UN News. The US and China haven't signed.
The week's theme
Every story today is about who gets to look. Resellers looking at prompts. Phone assistants looking at chats. Shopping agents looking at shelves. The companies that come out well are the ones that decided, on purpose, what each visitor may see. A seller can do the same on a smaller scale: let buyers and their agents see your products and prices easily, and keep your customers' private details in as few hands as possible.
That's the kind of setup we build at Vardast: one assistant that answers your customers on Instagram, WhatsApp, Telegram and your website in seconds, knows your catalogue, and passes the hard conversations to you. Ask us where your data goes. We'd rather you asked.