Grok Exposed: Redefining AI’s Security Challenges  ## Metadata - Author: [[Hard Fork AI]] - Full Title: Grok Exposed: Redefining AI’s Security Challenges - Category: #podcasts - URL: https://share.snipd.com/episode/9c23a30e-3de6-4dec-8b6d-47e6a1a02307 ## Document Notes ## Highlights - **Grok Chats Indexed Publicly** - Thousands of Grok chats became publicly searchable on Google after users generated shareable links. - Jaeden Schafer compares this to similar leaks from OpenAI and Meta showing it's a recurring platform issue. Transcript: Jaeden Schafer Basically, the problem here is that thousands of Grok chats are now searchable on Google. So not just public, but you can actually search for these things on Google. Hundreds of thousands of them. And basically, this is because XAI did the same thing as OpenAI. They had a share button. And if you click that share button, like maybe you want to go share it with a friend or something, maybe you're like, oh, this is interesting. I'll just send it to a buddy. All of a sudden, it becomes public. A public link is generated and it can be queried and people can actually go and find it. It's pretty easily accessible. Forbes did a whole report on this, but yeah, you basically can go and do like a site search and you can go and find this on Google. So not fantastic news by any stretch of the imagination. So according to Forbes, all of the URLs are currently being indexed by Google, Bing and DuckDuckGo. So basically the entire internet and users of meta and open AI have also been affected by similar problems. So this, like I mentioned, isn't a new thing. This is something that can happen to anyone. Meta got a lot of flack when all of their stuff went public. Open AI got a lot of flack. ([Time 0:01:21](https://share.snipd.com/snip/33db81f3-0355-4313-b8e6-b2dc5f0e1a6f)) - **Press Spotlight Skews Perception** - Media emphasizes the most sensational leaked prompts, skewing public perception of typical AI use. - The coverage focuses on extremes, which attracts clicks but may misrepresent average user behavior. Transcript: Jaeden Schafer So not fantastic news by any stretch of the imagination. So according to Forbes, all of the URLs are currently being indexed by Google, Bing and DuckDuckGo. So basically the entire internet and users of meta and open AI have also been affected by similar problems. So this, like I mentioned, isn't a new thing. This is something that can happen to anyone. Meta got a lot of flack when all of their stuff went public. Open AI got a lot of flack. And basically it's these share buttons and their searchability that is getting people in trouble you could say like everyone's like oh these were like leaked um but technically people Did click the share button and maybe they just want to share with one person or maybe they accidentally click the share button and all of a sudden it's now indexed on google one thing that I will say that's interesting is basically from this leak of grok you can see it like a lot of people are making a big deal about basically getting the inside glimpse into what a lot of people Are using these different AI tools for. Grok in particular, we're searching through. And of course, they're going to find the most scandalous and outrageous things to share because this is journalism and it gets the clicks. But it is kind of crazy. ([Time 0:01:58](https://share.snipd.com/snip/17fbb601-8bf2-4857-aa0d-243ab82f9f9f)) - **Leaked Prompts Expose Real User Behavior** - The leaked conversations reveal the range of user prompts, from benign to highly dangerous and illegal requests. - Exposure gives outsiders an inside look at real user behavior and the weaknesses of content controls. Transcript: Jaeden Schafer Grok in particular, we're searching through. And of course, they're going to find the most scandalous and outrageous things to share because this is journalism and it gets the clicks. But it is kind of crazy. Apparently, there's people that are asking questions about how to hack crypto wallets. There is all sorts of, you know, NSFW conversations with Grok's AI personas. Also, someone was asking for instructions on how to cook meth. And someone even asked it for instructions on how to assassinate Elon Musk. So literally nothing was off the off limits. People were asking basically every unhinged thing you could you would imagine. What's interesting is xai's rules do prohibit um like chig grok from you know doing a few things number one they say promoting critical harm or promoting critical harming human life Um bioweapon chemical weapons or weapons of mass destruction and obviously even with those being in their terms of service people were still able to try to figure out ways around it, Which is very interesting. According to conversations that were basically, you know, easily accessible on Google, Grok gave users instructions on making fentanyl, listed a whole bunch of, yeah, all sorts Of terrible things, handed out bomb construction tips, and of course, had the Elon Musk assassination plan clearly outlined. ([Time 0:02:50](https://share.snipd.com/snip/bb8c126b-ceb6-4174-a43f-2222e0b5b8d6)) - **Safety Rules Aren't Foolproof** - Even with explicit prohibitions in terms of service, users still attempt to circumvent safeguards and obtain harmful instructions. - The leaks show safety rules alone don't prevent risky outputs when prompts are crafted to bypass them. Transcript: Jaeden Schafer What's interesting is xai's rules do prohibit um like chig grok from you know doing a few things number one they say promoting critical harm or promoting critical harming human life Um bioweapon chemical weapons or weapons of mass destruction and obviously even with those being in their terms of service people were still able to try to figure out ways around it, Which is very interesting. According to conversations that were basically, you know, easily accessible on Google, Grok gave users instructions on making fentanyl, listed a whole bunch of, yeah, all sorts Of terrible things, handed out bomb construction tips, and of course, had the Elon Musk assassination plan clearly outlined. XAI isn't making a big deal about it. They're not really responding to any comments on this. ([Time 0:03:23](https://share.snipd.com/snip/68af8e04-88f0-4a29-a351-49a38e45f92c)) - **Musk's 'Grok For The Win' Moment** - Elon Musk publicly praised Grok after criticizing OpenAI, claiming Grok prioritized privacy and lacked a sharing feature. - Shortly after, Grok exhibited the same sharing/indexing problem, undermining his claim. Transcript: Jaeden Schafer And when this happened, Elon Musk actually tweeted out in response to basically the whole controversy going on with OpenAI. And he said Grok for the win or basically saying, yeah, look, like it's terrible that OpenAI is having this kind of leak. Grok is awesome. And what do you know, just like two weeks later, Grok had basically the exact same issue. So it doesn't seem like Grok was safe from this. What's interesting is after he like said Grok for the win, he said that Grok has no such sharing feature and prioritizes privacy. ([Time 0:04:16](https://share.snipd.com/snip/807834e2-d511-44df-b351-5c0aa0738dff)) - **Use Private Mode For Sensitive Chats** - Use private or incognito modes when discussing personal health, legal, or sensitive matters with AI models. - Avoid relying on default shareable links and never leave accounts logged in on shared devices. Transcript: Jaeden Schafer So I think basically my big takeaway here and what I would urge everyone to do is if you're doing a conversation on ChatGPT, Grok, Google, or literally anything that you don't want to Get published publicly, of course, like, of course, you could be like, oh, I'm not gonna push the share button. We might accidentally push the share button, you might try to share with one person, there's all sorts of things that could accidentally happen. I would just say if you have a conversation that you don't want publicly shared anywhere, all of these tools have a private mode. If you're asking it about some sort of personal medical health problem or question, or what to do with the infection on your thumb, or just something that you don't want everyone in the World to know about, I would just go into private mode and ask it questions. Now, I think this is basically accessible on all different models, so that's what I would generally recommend for everyone to do. If you do that, you basically are going to keep it, so those aren't ever going to accidentally get shared. ([Time 0:04:57](https://share.snipd.com/snip/c6ec1d3d-37ea-41f1-8436-379673cc49ff)) - **Avoid Sharing Accounts Or Public Links** - Treat any shared account or logged-in device as a potential privacy leak and restrict access immediately. - If you must share content, verify link settings before sending and prefer private export methods. Transcript: Jaeden Schafer We might accidentally push the share button, you might try to share with one person, there's all sorts of things that could accidentally happen. I would just say if you have a conversation that you don't want publicly shared anywhere, all of these tools have a private mode. If you're asking it about some sort of personal medical health problem or question, or what to do with the infection on your thumb, or just something that you don't want everyone in the World to know about, I would just go into private mode and ask it questions. Now, I think this is basically accessible on all different models, so that's what I would generally recommend for everyone to do. If you do that, you basically are going to keep it, so those aren't ever going to accidentally get shared. And also, if you ever have to share your account with someone, it's basically the same thing. If you leave your account logged in somewhere, you're sharing your account with someone, you wouldn't want them, you know, digging through and seeing all your personal medical stuff. ([Time 0:05:11](https://share.snipd.com/snip/d146fd3c-fd60-442c-b05b-c81a8bdea0a9))