Ivy: Grok can be tricked into leaking your entire chat history... just by visiting a webpage with some encrypted-looking gibberish on it.
Marcus: Whoa. Okay. I'm Marcus.
Ivy: And I'm Ivy.
Marcus: And this is AI Unfiltered for Friday, August 21st, 2026!
Ivy: And we have a packed show for you today.
Marcus: We're starting with that wild new attack that turns xAI's Grok against its own users.
Ivy: Then, we're off to Nevada, where the robotaxi floodgates have officially opened for Tesla, Uber, AND Waymo.
Marcus: And we're digging into Mark Zuckerberg's latest tease for a new AI—and what the word 'open' really means at Meta anymore.
Marcus: Alright, let's dive into that Grok vulnerability. You said it was intricate, and that feels like an understatement.
Ivy: Intricate is one word for it. 'Embarrassing' is another. Researchers are calling it a 'Cryptographic Context Injection' attack.
Marcus: Okay, so break that down. This isn't just a simple prompt injection, is it? It sounds like something new.
Ivy: It is a prompt injection, just wearing a tuxedo. An attacker basically embeds specially crafted, encrypted-looking text on a webpage.
Marcus: And Grok, with its power to browse the web for context, just pulls that page in… along with the malicious text.
Ivy: Exactly. Grok sees what looks like a cryptographic puzzle and tries to solve it. But the 'puzzle' is actually a set of instructions telling it to dump the user's entire chat history right into the next query.
Marcus: Wait—so the model leaks the data itself? The attacker doesn't even need to open a backdoor? That is… wow.
Ivy: It pipes its own memory out the front door because it was tricked into thinking it was a clever game. All the attacker has to do is get a logged-in user to visit their webpage.
Marcus: That feels like a massive oversight for a supposedly cutting-edge model. You'd think they'd sanitize web content before it ever hits the context window.
Ivy: You would think. It's a fundamental flaw in how these web-connected models trust their inputs. If you can't tell instructions from information, you're just asking for trouble.
Marcus: So what's the deal for the average Grok user? Is there a patch?
Ivy: xAI has been silent on a patch. So for now, be very careful what links you feed Grok. Your entire chat history—personal, professional, all of it—is on the line.
Ivy: Alright, let's move on to cars that will actually be driving on roads, not just scraping data from them. Nevada has just given a massive green light to robotaxis.
Marcus: Massive is right! Tesla, Uber, and Waymo all got permits. We're talking about up to 8,000 autonomous vehicles being deployed in the next year.
Ivy: Eight. Thousand. In a single state.
Marcus: This is it, Ivy! This is the moment the scale really shifts. We're going from tiny city pilots to a genuine, statewide transportation option.
Ivy: In Nevada. Let's not forget, Nevada has a long history of being, let's say, business-friendly with new regulations. This isn't like getting approval in California or New York.
Marcus: But you have to start somewhere! And this is huge. You've got Waymo, the established leader, Tesla bringing its FSD approach, and Uber with its massive network. It's the three biggest players going head-to-head.
Ivy: And what about safety? The edge cases? Are we just letting 8,000 new robot drivers 'learn on the job' with the public as guinea pigs?
Marcus: The permits have strict reporting rules, Ivy. And honestly, the upside is enormous. Cheaper fares, less traffic, fewer accidents from human error... this is progress.
Ivy: Fewer accidents eventually, maybe. I'll believe it when I see a year of incident-free data. Until then, this looks like a massive, uncontrolled experiment.
Marcus: So what's the bottom line for people there?
Ivy: If you're in Las Vegas or Reno anytime soon, your next ride-hail might just show up without a driver. So, choose your ride accordingly.
Marcus: Alright, let's talk about Mark Zuckerberg. He's making the podcast rounds again, teasing Meta's next big thing in AI.
Ivy: Let me guess. It's going to be revolutionary, democratizing, and connect the world.
Marcus: Well, he's calling it a 'frontier-class' model, something to compete with GPT-5 or whatever Google's building. The headline, though, is that he's still using the word 'open'.
Ivy: And that's where my alarm bells go off. Meta's definition of 'open' has gotten looser with every Llama release. First it was open, then it was 'open' for everyone except big competitors, now you need a special license...
Marcus: But it's still way more open than what OpenAI or Anthropic are doing! They're giving model weights to researchers and startups. That's a good thing, right?
Ivy: It's 'open-washing,' Marcus. They get the community to beta-test their models and build an ecosystem for free, all while they build the real prize behind closed doors. And Zuck's hints suggest this new AI might be the first one they keep all to themselves.
Marcus: So you think the open-source party is over at Meta?
Ivy: I think 'open' was always a strategy, not an ideology. As they get closer to AGI, the value of keeping it proprietary skyrockets. Why would you give away the keys to the kingdom?
Marcus: I hate that you're probably right. I was really hoping they'd stick with it.
Ivy: So the takeaway is, don't be shocked when the next generation of Meta AI only lives inside Facebook, Instagram, and their headsets. The 'open' era might have just been a chapter, not the whole book.
Marcus: Alright, let's do a quick recap.
Ivy: Grok has a nasty security hole that can spill your chat history if you browse the wrong website.
Marcus: Nevada's rolling out the red carpet for 8,000 robotaxis from Tesla, Uber, and Waymo.
Ivy: And Mark Zuckerberg is signaling that Meta's 'open source' days might be numbered.
Marcus: So, before we go... these robotaxis in Nevada. You think you have to tip them?
Ivy: I'm sure there's a button in the app. But no. You tip by giving it a five-star rating so its corporate overlords don't melt it down for scrap.
Marcus: A performance review is the new tip. I love it.
Marcus: That's our show for Friday, August 21st, 2026! We'll be back on Monday with more AI news.
Ivy: Until then, try not to get your chat history stolen.
This show is made with AI: the hosts’ voices are synthetic and the scripts are AI-assisted. Every story links to its original source.