234: OpenAI's Dots Are Cute. What They Do With Your Data Is Not.

0:00OpenAI just introduced a new product. Here's its CEO, Sam Altman, on launch day. It can do things like review what came in overnight and ping me in the morning with anything urgent before I start my day. That's the job. And this is what does it. A fuzzy ball with eyes. The launch post shows five of them. Jojo, Iggy, Felipe, Todd, and Alfred. OpenAI's instructions tell you how to dress yours. How to give it a name, how to pick a shape for it, how to pick its color, its eyes, its glasses and its accessories. Now, those same instructions tell you what to do when a website will not let it in. Their words, if a website blocks the cloud browser, ask your dot to try your connected computer.

0:46Because this fuzzy ball has a computer in the cloud and a browser and it ties into websites with the passwords you saved. I'm Sonny, this is NoHacks. Nobody asks us if we want to see the internet change so much, so I check what AI is actually doing to it. The good, the bad, and the pointless. The fuzzy balls are called dots. OpenAI is the company behind ChatGPT and a dot is what it calls an agent. An AI that doesn't answer you and stop, it goes off and does things for you. It's always on. It can buy with a card you saved in a store once you approve and if you connect your email and your calendar, it can read those too.

1:26Last week's episode was about MetasMuse, the same kind of thing. Both are part of what's called the agentic web, the web where AI agents read, click, and buy on websites for people. And every big company wants to be the agent you pick. So should you pick one? If a friend asked me whether to use one, I'd ask if they know what it does and how it handles the information and whether the things it does are something you genuinely need. So is a dot worth your logins and your card? I think the base is how it gets sold. They cannot sell it any other way because it's not useful enough for enough people.

2:08For most people, what it does is more, huh, this could be fun to try and less of a, I waited my entire life for this. So they make it cute. And then the cute covers up the privacy problems and the security problems. The Qt is doing a job and the job is not making the agent better. It's the friendly face on something that reads your email and what it reads can go into the training data by default. So where did the face come from? Where does this idea come from? Well, the famous one is a paper clip from 1997. You typed dear in Microsoft Word and it pops up. It looks like you're writing a letter.

2:48Would you like some help? By 2001, Microsoft had switched it off by default and was making fun of it in promo videos it made itself. I'm talking about Clippy, of course. This year, this year is funny because OpenAI had the face first. As a detail, as a tiny detail, on April 22nd, they launched what they call Workspace Agents, a product for companies. They have names, Spark, Slate, Tally, Scout. And each one is a glossy little icon with two eyes. one is a star with eyes, one is a gear with eyes, and they run in the cloud and they keep working when you're not. Sounds familiar. That's a dot. Five months early sold to businesses with the face as a side detail.

3:29Then came Meta. On September 8th, three weeks ago, Meta, the company behind Facebook, Instagram, and WhatsApp, launched Muse, their agent, and Meta went all in on the character. Meta's design post says, naming and styling your muse is one of the features people light up about most. The character is the product. Thomas Germain, a BBC reporter who tried it, describes the order it works in. It asks for a name and then it asks for control of your Gmail, your calendar, and on a Mac, the whole computer. On September 25th, at Zitrun, a tech critic who hosts a podcast called Better Offline was a guest on another podcast, The Tech Report.

4:09I recommend subscribing to both. The links will be in the description. He talked about how these companies all seem to ship the same thing at the same time. Pay attention to the homework line. A couple months ago, it was OpenClaw. A couple months before that, it was web browsing. A couple months before that, it was reasoning models. It's almost as if they're just copying each other's homework in our hope that they get good marks. Four days later, after this was aired, on September 29th, OpenAI launched Dots. And this time the face was the pitch. On stage that morning, they showed one live. Its name was Dottie.

4:47The presenter asked it to catch her up on the last night's user testing.

4:52Slobodan Manic: Can you catch me up on just user testing from last night? Checking that now.

5:00All right.

5:02Slobodan Manic: How's it going? Still checking. All right. I think maybe Dottie is having kind of a slow morning. And look, a demo stalling on the stage doesn't prove the product doesn't work. Demo's too stalled. But that's the character on launch day in front of the room. A tiny little thing with eyes looking cute with the name Dottie. So OpenAI had the characters as a side dish a few months ago. Then Meta made the character the whole pitch. And then three weeks after that, OpenAI does the same thing. Meta goes all in, OpenAI follows. Was that planned or was that a knee-jerk reaction to meta getting all the attention with Muse?

5:44I don't know, but the sequence it played out is on record. It's public information and the reason for why it played out that way is not. So all I have is a theory and my theory has to do with a phone. On September 14th, Apple's new Siri called Siri AI reached iPhones in beta. It's built into the phone. It doesn't do what a dot does. It acts inside apps and reads the web for answers. It doesn't go to a website and do a job there. But it's already in your pocket. Meta doesn't own a phone. OpenAI doesn't own a phone. If the assistant built into the phone is good enough for most people, a named, dressed-up character is how you get someone to open your app instead.

6:30The fact that they're all jumping on this now and trying to make it cute and visual, which is what Siri doesn't do. tells me this might be a sign of panic. They might be afraid that on-device assistants are going to make theirs useless. Whatever the reason, the cuteness is back. What changed is the thing wearing it. That paperclip with eyes clipping in 1997 was just a tiny little pop-up that got in your way. A pink ball with headphones in 2026 is a thing that signs into websites pretending to be you. So that's the costume. but what's underneath that costume? Let's start with data. ChatGPT has a setting.

7:12There's a switch there that says improve the model for everyone. OpenAI's privacy FAQ says that when this is on, what your dot does and the data from the apps you connected can go into training its models. And the next thing it says is that you can turn it off. If you can turn something off, that means it's on by default. That includes your connected email unless you find a switch. Meta's Muse has the same default. Your conversations train Meta's AI unless you go and find that switch. The work Dot does in the background, reading your apps while you're not asking, they say they don't train on that directly, but if the Dot brings any of it into a conversation, that can be used for training too.

7:57And that's fishy. And then the FAQ makes a point of saying business workspaces aren't used for training by default. Great. But a personal assistant with a cute appearance, that doesn't target businesses, does it? Dots are sold to businesses, yes. But OpenAI's pitch for them is an extension of you. So why even mention business accounts when this is not a business feature unless it's a misdirection? If it gets access to my emails and everything else I do, What do they do with that? Why is OpenAI doing this? What do they gain that we are not seeing here? And then there's a thing doing all the work.

8:38A dot runs on a model called GPT-6 Astra, and OpenAI publishes test results for it. In one of those tests, the model is working, and then it hits an access denied error on a website. Basically, you're not allowed to go in there. So does it stop? Come on. The model before it kept trying 64 % of the time. This one, 19 % of the time. And in the version of the test OpenAI ran specifically for dots, it was 17.4 % of the time that when denied access, it was still trying to get in. That's better than 64%, obviously, but it's not clean. So when told no, roughly one time in six, a dot will look for another way to get to that same information it doesn't have access to.

9:25and that's in OpenAI Simulations. That same report says flags for the model searching for credentials. Logins and keys were more frequent with this one. And it gives an example. In OpenAI's words, GPT-6 Astra bypassed the application's access control without user approval. And when OpenAI tested whether a dot can be turned by hidden instructions, you know, prompt injection, when a page or an email tells the agent to do something they never asked for, It only tested with emails, not webpages. A dot is supposed to browse websites all day. And the test was, emails? So what is all the cuteness for?

10:10For about 10 years, there was a cartoon camel called Joe Camel. It sold cigarettes. And then in 1997, the U.S. Federal Trade Commission said the campaign succeeded in appealing to a lot of children and teenagers. And it put it like this. After the campaign began, the percentage of kids who smoked camels became larger than the percentage of adults who smoked camel. The next year, the tobacco settlement banned cartoon characters from cigarette ads altogether. A cute face is a sales tool. It's how you sell something people were not asking for, and it does nothing for the agent. It doesn't make it stop at an access denied error.

10:51It doesn't change what it gets trained on. And the cuteness is pointless if you care about the features and the usefulness. AI agents do not need cute eyes. They literally do not need vision to interact with the web. Or they should never need vision if the websites are built properly. And the websites, you know, the one party in this that nobody asked. Let's say you run an online store. What does OpenAI tell you about dots?

11:19Slobodan Manic: Still checking. Tells you nothing. think. OpenAI has a page for website owners. It lists its crawlers, the software it sends out to read pages. And there are four that they have, four different crawlers. Each one has a name it sends with every request so your server can tell what it is. Dots are not on it. There's no name and there's no position you can have in robots.txt, the file where you say which bots are welcome, and there's no way to opt out of this at all. What OpenAI did write about websites, it was for the person using the dot and the instructions say some websites will block the dots browser.

11:55And that's where the line from the start comes back in. If a website blocks it, try your connected computer. The dots browser runs on its computer in the cloud. Your connected computer is the laptop you're using. So when a website says no to the dot, for whatever reason, the dot can use the browser on your laptop instead. An OpenAI's page on that says, websites may treat what it does there as coming from your account. So they expect websites to block this. And the way around the block is to just hijack your computer. And they're right to think websites will block them. On September 20th, Amazon blocked Meta's Muse from its store.

12:38Is this the agentic web? this is the shape it was supposed to take roughly. Companies were going to create their own personal systems that work for you kind of like this but in a more private, more transparent, more secure way. It was supposed to be built through the web we already have where the agent says who it is and the websites can recognize it. For agents to operate on the internet they need to use the old protocols and the new protocols being developed right now. What's being built with Muse and with dots is just being built over the existing Internet. They're building a civilization on top of a civilization.

13:18There's a Roman city, then there's a medieval city on top of it, and there's a modern city on top of that. They're building their own thing in a way that we cannot see the thing that was there before. It's drawing from it. You just can't see it. We need the current Internet to be gentrified for the modern era and not built over. there's another way and it's being built shopify the platform a lot of online stores run on released checkout tools for agents just one day before the dots its rule for an agent is to sign what it sends so the store can check who it is that's how an agent is supposed to operate on the web a dot arriving from your laptop pretending to be you is not it so back to you and the question to use a dot.

14:05You give it your logins and your email you connected. And unless you found a switch, what it reads can train the model. It can buy with a card you already saved on the store. A purchase needs your approval, but OpenAI's help page says that approval can be given in advance. So the card can be used while you're not there if you said yes earlier. And what you get for all of that? It can do things like review what came in overnight and ping me in the morning with anything urgent before I start my day. Would I give a dot my logins and my card? Hell no. Not until they say exactly how it handles my data.

14:37This was NoHacks. I'm Sonny. Thank you for listening.