Stripe Buys OpenRouter for $7B, New Models Drop & Anthropic Watermarks AI | This Week In AI

Shane: The joke is that the GitHub status page looks like a Christmas tree.

Abhi: All of us were saying that, oh yeah, the AI generated code is the reason or whatever. But then now it's like getting to a point where is it a discipline problem? Is it a team problem? Or talent is leaving.

Shane: Alright everyone, welcome to Agents Hour. If you're just tuning in, we're gonna do the news. It is Monday, August 17th. There's been a lot going on, there's a lot to cover. We do this every week. If you're tuning in after the fact, whether you're watching on Apple Podcasts or Spotify or anywhere else, YouTube, thanks for watching. Give us that subscribe. Give us that five star review. That helps other people like yourself. Find the show. Or you could just tell a friend. But a lot of times people tell us they don't have any friends.

Abhi: Their agents are their friends.

Shane: Their agents are their friends. So if that's you, then The review helps. Otherwise, I'm sure you have some friends that might be interested in the show.

Abhi: I'll be your friend.

Shane: Yeah.

Abhi: Please If you give me a five-star review.

Shane: Yeah. Friendship comes cheap.

Abhi: Dude, transactional friendships in San Francisco. It's that easy. Five stars.

Shane: Let's start off with something a little more fun. This was something that I saw pop up on my feed. I just found out my best friend broke up with their SF boyfriend over the weekend. I asked what happened. She goes, I found out he had an AI agent managing the relationship. Turns out the thoughtful texts, the check-ins, the good morning messages, the conflict resolutions, dinner date plans are all were automated. My friend was emotionally attached to a cron job.

Abhi: Dude. Honestly, it's beautiful.

Shane: This is the world we live in now. The world we live in. The sad thing is, this is what's going on in SF. But if it's going on in SF now. In six months or twelve months, ladies, it's gonna be everywhere. So be prepared.

Abhi: If you've got a message me from me on Bumble, you know that it was not me. At all.

Shane: You know, it's kind of a wild world that yeah. Because I think it probably starts with just asking the agent for advice.

Abhi: Yeah.

Shane: And then you cut out the middleman, which is yourself. You just let the agent go direct.

Abhi: Then you train it on your voice. We do this for Twitter posting. So like you can imagine people can do this for text. Your agent could live an iMessage now, thanks to different providers. You could totally pull this off.

Shane: Yes. I mean it even you can pull it off. It doesn't surprise me actually. Alright, let's talk about Anthropic watermarks. So just in Claude models will now have invisible watermarks embedded in all text and all metadata attached to files. I think this originally started as kind of maybe it was EU regulation or requests from EU. And then Claude basically came out or Anthropic came out and said, you know, our content, AI generated content is gonna have watermarks. The way that I understand it works, it's not necessarily just metadata. It detects like the words and then it can give you some kind of accuracy of or percentage of yes this is most likely AI generated because these watermarks are here in this certain pattern. But also I'm paying for you to generate it. I also maybe I don't want it watermarked. But I think I think that's a lot of people are kinda giving some backlash around that. Because even if you do take out parts, revise it, those watermarks probably still persist to some level. In some ways I think it's okay, but in other ways, what if ninety percent of it was me massaging the output and using it for just generation, but the watermark persists? Like at what level. Should it be watermarked or shouldn't it be? What if I just use it for an outline? I think this has the potential to maybe push more people to using open models if they don't. Do the same type of hidden watermarking process. I think Anthropic was also gonna have an API so you could run content through it to figure out if it had traces of this watermark thing in there. How does it work with code?

Abhi: You know, these are questions I have.

Shane: I haven't read into it enough, but these are interesting times because I think we're gonna see this from more people.

Abhi: And of course the EU is the first people to demand these.

Shane: And I bet Anthropic just said, well, if we're gonna have to do it for the EU, we can't just like make a change just for the EU, so we'll have to make it for everybody.

Abhi: But I also wonder what does that mean for OpenAI? Are they gonna follow suit? Well I'm sure. I mean I think all the Frontier Labs that are operating in Europe will have to do this Transparency Act or whatever it's called. It was a EU AI Act, Article 50 Code of Practice on Transparency of AI generated content.

Shane: Most of the time when I'm looking at AI generated writing, I can tell. You know, it's usually not great. But ultimately if the writing is good, I don't care if it's human written or AI written. I just can usually tell that AI written writing sucks. So You probably just shouldn't do it.

Abhi: I think over the weekend a lot of people got roasted for M dashes again, so you know, that's become a meme. I don't know. I don't know why we Should care too much, right? Yeah, as long as the content is good.

Shane: But I could see, you know, especially with videos, maybe it's it matters more. You know, you don't want to represent it looks like someone's in a video. Like there are some real reasons why having watermarks could be good, but I think we will find out more. This came out not exactly related, but it is kind of related. So they found a way to and you can see this got 3.3 million views. So a lot of people thought this was pretty interesting. Found a way to extract hidden reasoning of frontier models using a vulnerability in the APIs of every frontier AI company. They verified the reasoning token count matches build API thinking tokens one for one for most of the prompts queried. So back in May, they found that encrypted reasoning could be replayed outside of its original context. They reported it to the labs and they said they don't see security implications. So This means that like Haiku 4.5 can read Opus 48's thoughts if you do a little bit of jailbreaking basically. This suggests that distilling reasoning traces may have been possible for a long time without ever breaking the cryptography. Some other things that came of this, the big one is that if you ever shared an online Claude Code or Codex session with encrypted reasoning blobs, they can be decoded and leak your personal data. So there's just a lot of interesting things. You can read the thread. Yeah, you can read the paper about it as well. What do you think of this?

Abhi: I think it's very surprising that they don't think that's a security hole, you know. Maybe they would if like someone could successfully distill from it, then maybe Dario would be all pissed. But until that's proven I guess, you know, it's not that big a deal. But leaking personal data is always a big deal, so Yeah, but I think it's like you gotta have you'd have to have the link.

Shane: But I do think that I saw other p reports of the links could be searchable in certain context as well.

Abhi: So if it's like a public converse.

Shane: If you share it, it becomes a public link. And it's you know, you need to know the URL, but Yeah, but you also have to log in

Abhi: So I guess like the threat vector is like you have to jump so there's some hoops. Normal fools won't do that.

Shane: We had a bunch of new models over the week. Grok 4.6 came out. This was from August 12th. It delivers Frontier Intelligence and is a significant improvement over Grok 4.5 at the same price. You can see how it compares to Grok 4.5, how it compares to 5.6 Sol, how it compares to Fable. And it is right there. You know, some benchmarks it's ahead, some benchmarks it's just slightly behind, but it basically puts itself in the conversation with 5.6 and fable.

Abhi: Yeah. I daily Grok last week. So great daily 4.5. Didn't necessarily like it, but I promised Evan I would. I did that. Four six came out, tried it, and then I was like, oh, it's Friday. Don't have to do this anymore. But maybe I'll do it. Did it pass the vibes or was it passed the vibes, but I didn't you know, at that point I was like I felt like I was working handicapped last week. Because it's not it's just not as good as like 5.6. But then 4.6 was really good. I would like I have this test I do with new models now, which I just say make bomberman the game. Like a Game Boy version of it and it didn't one-shot it but it did pretty good job. So I was like I was pretty impressed but I haven't done that same test on 4.5 because I was just doing that For work, work, you know.

Shane: And then Elon came out that day and said Grok 4.7 is better than 4.6 and should be ready in three to four weeks. The idea is we just got 4.6 and there's already something in the works that's gonna be a lot better. Could xAI leapfrog, could Grok leapfrog the other two models?

Abhi: I'm starting to think so. Mainly because we have friends there now, and all they're saying is frontier soon. That's all it's they can't reveal anything. They're just saying frontier soon.

Shane: For a long time I kind of thought just it was it was Anthropic and OpenAI's race. And now you have Grok. Pretty competitive. You have meta, which is they're in the game.

Abhi: They're still kind of a little bit behind, but they're in the game.

Shane: Especially if they're leaning more towards open models. Who would have thought that, you know? Meta's gonna be the open champions, you know. You kind of thought it back when Llama was a thing and then it kind of crashed. But now maybe they're-emerging and they're acting we're gonna have some competitive open models from the US. So Game is still open.

Abhi: I mean I guess we could check with CodeRabbit after on how they benchmarked, because they benchmark models all the time for their stuff. I'm curious how 4.5 or 4.6.

Shane: Speaking of open models, GLM 5.3 was released on August 14th. It says Built to code, ready for cyber defense. I've been seeing a lot of hype around how good it is for cyber defense. But you can see it, you can look at the benchmarks here. It's better than five two by quite a big margin on almost every benchmark. Still behind, you know. Fable 5 and 5.6 on most things, although it's very close on a couple things. I mean, and actually better on a few benchmarks. So it's near frontier level intelligence. Yeah. And I don't know if they've opened it yet or not, I guess, but I missed it.

Abhi: Not on Alibaba Cloud yet. It's not. I started thinking about this too. Like I want to try this out, but I have to wait. Which I guess m pushes most people to go get a Z plan. But then I'm like, Grok is like as good and cheap too, so maybe I w don't even care.

Shane: Yeah. I mean I think people do care if it's you know, once it's open, because then There's always the even though no one's running this model, right? No one's running this model themselves, you know, but it's the idea that you could, right? Or you can use one of the other providers and just pay, you know. Basically pay inference for it. This is from Qwen. It says we promised open weights for Qwen3.8. Now time to meet them. So they have Qwen3.8-27B. And so this is impressive because it's a model you could actually feasibly run yourself. I think I saw it's nearly like Opus 4.5 level on a lot of tasks, and it's b basically you can run it yourself. There might be. I'm pretty sure it was Opus 4.5 when I when I was looking into this, that it beat Opus or is close to Opus 4.5 and now it's a 27B model you can run on your own hardware, which is wild.

Abhi: And then Qwen3.8-Max is on Alibaba Cloud. So I'm about to burn some tokens.

Shane: Yeah, and then you have the 3.8, a 2.5 trillion parameter model, which is now open weights. And so now you can, you know. Test that out too. And then yeah, this is just talking about the 27B model. I think there's been a lot of hype around local inference. Just imagine was that maybe it was a year ago, whenever 4.5 came out and it was amazing. Then imagine fast forwarding saying you can actually have that. Or close to that on your own device.

Abhi: Yeah.

Shane: That's nuts.

Abhi: I mean think about like, okay, four or five was a great model, right? And sure, like, is it not as good as maybe it's not as good as what we have in the frontier right now. Everyone was hyped about that. Now you can have it for free if you have the machinery. Or just pay fraction of the price to have that same fun. You know, maybe people want to rebuild GitHub and stuff, like you know, you can go do that.

Shane: Yeah, just let it run.

Abhi: For cheap.

Shane: It's funny because we were j earlier we were talking about how a lot of the compute and a lot of your coding is moving to the cloud. But I still think there's this idea of I just want to pay electricity and I want to run something for days on my own computer because it's free. Right. And so I think the tooling to run things locally is getting better. I think the models are getting better. And again, I'm not saying this is four or five. I haven't tested it. It's just kind of what I've heard that I think some of the benchmarks, it's pretty close to four or five.

Abhi: And it's wild that you can just run that locally. Like, share, and subscribe.

Shane: And follow us on X.

Abhi: And tell your friends. And their friends. I mean, we're not begging.

Shane: Well,.

Abhi: Maybe a little bit. Subscribe to Agents Hour every Monday, noon Pacific.

Shane: There's a lot happening. Honestly, the last couple weeks. Money's moving. There's a lot of consolidation as far as like acquisitions and also a lot of fundraising going on. So this was a huge one. Stripe has finalized an agreement to acquire OpenRouter. A startup that helps companies switch between artificial intelligence models for more than seven billion according to people familiar with the matter.

Abhi: Yeah, dude. They just raised 1.5 billion valuations.

Shane: So they're basically 5x in three months.

Abhi: Which is wild.

Shane: And yeah, and we've met Alex before. We know a few people over there. Congrats to them. This is an awesome outcome. Your growth has been tremendous. I mean, their revenue ramp, their token ramp has been. Insane.

Abhi: Another one of Sam's fraternity bros is rich.

Shane: So yeah, that's awesome. Like an amazing outcome. Now let's talk about like why.

Abhi: Yeah, why?

Shane: Why is OpenRouter that's the question everyone's asking, why is OpenRouter worth seven billion dollars? Make it make sense.

Abhi: There's a premium there going on for sure.

Shane: Typically for fifty times reported revenue, unless you think there's either insane growth or there's some kind of extreme strategic value that it would be providing Stripe.

Abhi: And the technology.

Shane: So I have I have my theory. What's your theory?

Abhi: I think our theory is probably the same. But I think it's more about the technology and Stripe being a player in this game since before this era. They got money. That's number one. They have money to blow. Do they want to go recruit a bunch of Gen Z kids in SF to build this product themselves? Or do you want to go get this solid team that is making money? With solid infrastructure and technology and just grab 'em and you know and then you have to pay for that.

Shane: Yeah. And we have no inside information at all on this, but I think. I agree with that. I think we had a conversation a while back about r Ramp becoming an AI company. Ramp released their gateway. I feel like Stripe and Ramp are on this collision course where they're both kind of in the finance financial sector, but they both want to really become known as AI companies as well. And I think Stripe saw what Ramp was doing. Stripe's been, you know, releasing a lot of information on their blog around their developers becoming more productive. They're moving to cloud agents. They're like

Abhi: They've had this like weird competition, I think, even though they're not they haven't talked shit about each other yet.

Shane: Yeah, they're not they're not exactly overlap, but I think they're like heading in the same direction. There's a collision course. I think Stripe just said. You know what? It's gonna take us six months to build this ourselves and we're still not. Yeah, and we're likely not gonna get that kind of traction. We'll just pay a premium. And then now we've just inserted ourselves in the game.

Abhi: Yeah.

Shane: Now they're firmly in the game and we'll see what they do next.

Abhi: Yeah. Like tokens are a form of currency now. So at least for agents.

Shane: Yeah, you want to be on the token chain, right?

Abhi: You're on the inference chain in some level.

Shane: And if you can be the one that's routing between models and you know, I'm sure Stripe has some. Yeah. They've been doing a posting a lot about how their developers are using it, so maybe they see the perceived value. I'm sure a lot of developers at Stripe were huge. Users of OpenRouter.

Abhi: Imagine you start your business, you do your business registration through Stripe. You get tokens just given to you for signing up. And then if they do expense management and all that stuff, which I'm betting they're gonna do the whole life cycle, you could run your whole business on Stripe. You don't need a Brex, you don't need ramp or anything, you know. So I would say like everyone else should be scared too.

Shane: All right. Another friend, Aparna, had this post. She's from Arize. Jason and I started Arize six plus years ago with a simple proposition that headlined our seed deck. We make the world's AI work. Today we're announcing we've entered into a definitive agreement to be acquired by Dynatrace to accelerate that vision. So Another acquisition. Congrats to friends at Arize.

Abhi: Billy, dude.

Shane: Yeah. That's been another, you know, good outcome, great multiple on, you know. Like just what how they've been growing. I always hear about just how good they are for enterprise, right? Like that's where I always hear Arize come up is a lot of the big users of Mastra or others that we talk to just really see Arize as kind of the enterprise way to get observability on the traces. So definitely. Interesting to see how fast they've grown over the last couple of years and congrats to them and then the whole team.

Abhi: And like Dynatrace was like founded in like 2005. So they've been around for like twenty-one years. They're an observability company. Now they're gonna be an AI they are an AI observability company. Arize is really good for AI observability.

Shane: And Arize has Phoenix, which is like their open source version as well.

Abhi: So they have the evals and all the stuff, so I think it makes sense. Yeah, I kind of forgot about Dynatrace.

Shane: So maybe is there a way to, you know To get back in the game.

Abhi: Get in the game. Yeah, they're from Boston. You know. People don't know who Dynatrace is, let's be honest.

Shane: Yeah.

Abhi: But now they'll know because of Arize. And I guess that's the big move they wanted to make, right?

Shane: Because they're public, right?

Abhi: They're public company.

Shane: We are in the CodeRabbit office and A big announcement from CodeRabbit happened last week. This is August twelfth. It says we raised a one hundred and forty-three million dollar Series C at a 1.5 billion valuation to build the control layer for software change. Our first investor got involved for a totally different reason. And so we gotta watch this video because why not? For those of you that have ever watched Silicon Valley, you'll appreciate it.

CodeRabbit Video: Hey Josh! Thanks for seeing me today. Take a seat. Oh. Sick. So, as I'm sure you already know, your portfolio is through the roof since the $143 million round CodeRabbit closed. CodeRabbit? Yeah, you invested heavily and early. You were the lead investor in the CDN Series A round. Yeah. I remember thinking it sounded like there was like a rabbit coding and I wanted to be part of that Wow. Well you certainly were because now you're a certified billionaire. Oh wait. What does CodeRabbit do? Let's take a look. Build code base aware context across pull requests, repositories, and developer workflows to detect defects, security risks. Oh. Do you mind if I Yeah, no, go ahead. Let's go. Okay. Security risks and architectural regressions, then generate actionable fixes before merge. For sure. So where does the rabbit come in? There's no rabbit. What? Then what's all the money for?

Abhi: On point.

Shane: So that was hilarious. That was a that was fun to see.

Abhi: So playing Big Head, dude, that's dope.

Shane: So congrats to friends over at CodeRabbit. That's a big deal.

Abhi: Dude, do you think if we raise RB or C or whatever, we can get Dinesh? We just keep this story going. Because Railway had Gilfoyle.

Shane: Gilfoyle, yeah.

Abhi: And then now that now Big Head’s on the table.

Shane: I feel like Dinesh would be the hardest to get.

Abhi: Hardest because he's like a movie star.

Shane: Yeah.

Abhi: But I don't want Richard though. He's whack. Maybe we'll get Erlich. Erlich. We'll get Erlich. TJ. We'll get TJ Miller.

Shane: Lovable. Also announced a raise. So they raised $400 million at a $13. 3 billion valuation to help you achieve more than you ever thought you were capable of. So lovable keeps growing. That's a huge, huge, huge raise. Congrats to the team over at Lovable. When's the last time you use Lovable?

Abhi: It's been a while, dude.

Shane: Yeah, I mean I was a huge user about eighteen months ago.

Abhi: I just use Replit now, honestly.

Shane: Yeah, I mean. I use Replit or I just I feel like if you already are running your coding edge locally or in the cloud, you just do it there. But I do know of So I have a friend who has used Replit, uses Lovable, uses them both, you know, depending I think kind of bounces back and forth, and he's less technical than me and he loves tools like this.

Abhi: So there's still a huge market.

Shane: Of folks who don't have any understanding or a deep understanding of how writing code works, how deploying code works, and if you can do it all in an all-in-one platform. And love lovable does have a lot of, and same with Replit, a lot of connectors to make it really easy. So congrats. All right, let's talk about agent harnesses and infra. So DeepSeek Harness v0.1 is now available in developer preview. So it's opened up. Is it TypeScript based?

Abhi: It's TypeScript, yeah. Looks very much like Pi.

Shane: Yep, very similar to that. So if you're a DeepSeek fan, try that out. Cerebras announced on August 13th, previewing ultra fast mode for OpenAI's GPT-5.6 Sol powered by Cerebras. Generates responses at up to 750 tokens per second. Just wild for a frontier model of that size.

Abhi: As a stockholder in Cerebras, I'm very happy because we are up fifteen percent. Not financial advice though.

Shane: If you're watching the show, you should know that anything we say on here is definitely not financial advice.

Abhi: Not advice at all. But I'm a Cerebras stockholder. I'm just gonna let you guys know that

Shane: Alright, we gotta talk about GitHub.

Abhi: Alright, alright.

Shane: So before we go into this, GitHub was down today.

Abhi: It was down last week too.

Shane: So the joke is that the GitHub status page looks like a Christmas tree.

Abhi: Yep.

Shane: Because it's should just be all green. And it is definitely not.

Abhi: It's a lot of red.

Shane: And you know, so much so that like people track the real actual GitHub downtime because they d. Their actual status page doesn't always show all the issues, so they h hide some of the issues. And that looks definitely even more like a Christmas tree. So what's going on at GitHub?

Abhi: I don't know man. Like we I guess all of us were saying that oh yeah the AI generated code is the reason or whatever. But then now it's like getting to a point where is it a discipline problem? Is it a team problem? Usually you go through a bunch of retrospectives on outages and stuff and you try to not make the same mistake again. Let's assume that these are all new mistakes, right? That means you're just waiting for the next mistake to happen versus being defensive about your architecture and infrastructure. Just because you like solved one fix, you sort you should start thinking about what are the other things that could go wrong. So I was saying this earlier, I think there must need to be a change in leadership. Or talent is leaving. One of the two.

Shane: Yeah, so I agree with most of that, but I'll play the devil's advocate. GitHub probably is getting hammered at a scale. That they could never have predicted two years ago. And so if you think about the sheer volume of because it's just agents writing, I mean if I bet if we looked. You know, and I know they've shared some numbers of the number of repos, the number of people h interacting with their APIs, probably a scale that they have not been built for. Honestly. Like you couldn't have predicted it. And so as much as I You know, fault GitHub for not being up because so many people depend on them. I can probably see how they got there, right? And they're true probably scrambling to try to make the infra scale. Using coding agents to probably help them get there, making mistakes along the way, because if they don't, they're gonna have downtime for other reasons. So I do think unfortunately for GitHub. I just questioned maybe the info wasn't built for scale that agents are gonna drive, and that does open the opportunity for someone else.

Abhi: Yeah.

Shane: Because someone else isn't gonna hit the same scale issues right away, right? You can slowly and if you build it in and architect it for agents for that level of scale from the beginning you might be able to get there.

Abhi: But you would think, right, you're a Microsoft company, you have unlimited resources, theoretically. Why can't you iterate? Is it because the code base is over twenty years old and it's hard to do?

Shane: Maybe.

Abhi: Is it because that every product owner that released every feature in GitHub is now working somewhere else and now you just have a bunch of fools? Who don't know the history of the code base, most likely, you know.

Shane: It's po code base is probably massive and make questionable how well even the frontier models can work across that.

Abhi: Also another thing is GitHub a sexy company to work for is the best talent. Trying to work at GitHub?

Shane: I feel like GitHub is almost like working at like the utility company.

Abhi: Yeah.

Shane: It's like a core system of anything that we do, but It's not the sexiest company to work for anymore. At one point it probably was.

Abhi: Especially if you can go work for Cursor or you can go work for these other places. It's kind of a shame, man. That's all I gotta say. Like last week we had actions outage. Today a lot of the services were out. They're still being investigated. But there's the opportunity for other people today. There are lots of opportunity today for others.

Shane: So let's talk about this. This came out today. This is from Cursor. Origin, our code hosting platform, is now live.

Abhi: Very timely release.

Shane: It's fast, easy to use, and deeply integrated with Cursor. Get started by syncing your repos from GitHub.

Abhi: Cool.

Shane: Yeah. Friend of the show, Matt Biilmann from Netlify, said Netlify is now a Git host. And they basically compared speed-wise to GitHub on a lot of things. It is faster. But what are they calling it?

Abhi: Netlify source. Netlify source.

Shane: So Netlify's trying to get in the game as well, being a Git host.

Abhi: Yeah. I suspect any info provider is gonna get in the game. Depending on how the revenue goes for Cursor Origin as well as like Netlify or whatever.

Shane: It's such an uphill battle to play though. Even if Netlify source is great. Benchmark-wise, it looks pretty good. But if you had other things that weren't hosted on Netlify, would you want to use Netlify source as your Git source? I mean, I guess you could. There's nothing stopping you.

Abhi: Yeah, but why would you, right?

Shane: Yeah. I mean I don't know. I feel like it really there really shouldn't be that much lock in on GitHub because you just pointed out a new origin and it should just all work.

Abhi: Yeah.

Shane: But there's a lot of things that probably won't, right? Like you have GitHub Actions and all these like these things that you set that you've set up. Yeah, you've set up over time. And now you probably have to rebuild all that in probably a different way. And so there I think there's some friction. But maybe with agents it's a little easier than it used to be. Swap.

Abhi: Yeah. Like I also wonder, like origin just came out, like the scale of that will scale to however many cursor users want to use it. But it's not the same scale as GitHub. And so why do we believe that these new products are gonna scale better than GitHub?

Shane: I don't think they will, but if they're if they have some kind of slight sexy factor that gets people to want to do it. Especially small teams that just getting started or that have a project that is more easier to port over. It's one of those things that maybe builds up over time, but I think it's pretty uphill battle. Alright, we got a bunch of quick hits. So new from Ramp AI Index, disappointing adoption of Fable 5. So mostly people just saying Fable 5 is too expensive. If you do compare just GPT-5.6 to Fable 5 and now you compare Grok 4.6, Fable is just significantly more expensive. And so maybe not everyone will pay for what's perceived as just a very slight edge. On most things. OpenVuln is a public vulnerability intelligence platform for open source. So if you run an open source project, you can submit your public GitHub repo and They have an AI engine that'll scan it for vulnerabilities and it's audited by GLM. This I thought was pretty funny. So a man representing himself in a Connecticut court hid prompt injections in official filings designed to manipulate AI. So he basically knew or predicted the lawyers on the other side would use AI.

Abhi: So funny.

Shane: And so he put pr hidden text in the filings to try to prompt inject the

Abhi: This is why we need watermarks, dude.

Shane: This is why. This is from Linear and this is from last week. It says it's the first week in Linear's history when issues created through MCP surpassed application created ones. So MCP usage is through the roof. Wow. So people are not using UI as much. They just want their agent to be able to do everything. Yeah. Makes sense. And that's it. Did we miss anything? Any news items we missed this week? Drop them in. But yeah, it was a good show. Great show. It's just wild to see. All the acquisitions, fundraises, there's just a lot of money and interest and I feel like things are only just starting to heat up. I think we're gonna see a lot more of this over the next couple of months.

Abhi: You know what's interesting? I had to keep that series C announcement from CodeRabbit a secret and I I wanted to tell everybody. I mean I didn't sign an N D A. I just did like, you know, Scout's Honor.

Shane: Friend DA.

Abhi: Friend DA. But yeah, super happy for them and impressed. Really want one of these paintings for our office too. Like I need a Renaissance mural. That's that'd be sick.

Shane: Yeah, is it what you need? A series B to get that?

Abhi: I don't know. I think you gotta be cooler than we. You gotta get it from Gamut. Yeah, Gamet. Commissioned and produced by Autonomous Agents, built in minutes.

Shane: Gamut. Open invitation. Come on the show. Paint us. Get us a painting of Me, Abhi, and Sam, you know, we'll get you on the show. For those of you tuning in, we appreciate you. Give us that five-star review on Apple Podcasts, Spotify, whatever. Subscribe, like, do the thing on YouTube. You can see. You can follow me on X @smthomas3, follow @abhiaiyer. If you know of someone we should be talking to, or you know, have them DM us. DMs are open. We're always looking for interesting guests to talk about what we're building with AI. That's it for this time on Agents Hour. We'll see you next week. Peace from CodeRabbit.