Transcription
What a wild day at Enthropic, cuz Claude Code's source code just got accidentally leaked, revealing not only internal systems, but also hints at future plans as well as new models.
Over 500k lines of code were exposed in a faulty update, and it spread across the internet within minutes. Before Enthropic could contain it, the code was already downloaded, mirrored, and impossible to fully remove. Things even escalated further when a developer recreated the system from scratch and released it publicly on GitHub, making the situation permanent and essentially making Claude code fully open source.
But the craziest part isn't just the leak, it's what's inside. The files reference upcoming systems like Claude Mythos, the new Kappa Barrett here. There's also hints of the Opus 4.7 and Sonnet 4.8 being tested internally along with new tools to Claude Code as well as a new whole Enthropic tool called Claude Buddy that most people didn't even know existed. In short, this wasn't just a leak. It's a full-on glimpse into Enthropic's road map. And it's something that we're going to be uncovering throughout today's video.
So, to get started, let's take a look at what happened with this leak. Essentially, a developer around 4 a.m. accidentally shipped a debugging file in GitHub and that essentially exposed over 500k lines of internal source code. It was discovered rapidly and it spread online within minutes. And despite takedown attempts, it was quickly mirrored and became impossible to fully contain. And it is something that exposes internal model code names and 44 hidden feature flags giving a rare look at Enthropic's entire agent system.
This is something that included different lines of TypeScript revealing how cloud code is built with a full agent architecture. You have also strict permission controls, token optimized designs, and a lot more. It also uncovered code names like models that we had actually covered previously on this channel like Fenick which is the Opus model. Then you have Tangu which is Haiku. Then you have Capra which is the Sonic model and Titan along with hints at unreleased versions like a 1 million context Capara model.
On top of that, there's 44 feature flags revealed with upcoming capabilities coming to Cloud Code as well as the Enthropic ecosystem. You have a background AI agent called Karios, a companion style AI called Buddy, a voice mode, multi-agent coordination, and advanced planning tools. There were also systems like an undercover mode to hide AI identity, and a structured long-term memory system, which is going to fix Claude Code's memory problem.
Before we get started, allow me to introduce today's video sponsor and a longtime supporter of the channel, Scribba. It's the only coding platform I found where the video and code editor are combined into one. That means with Scribba, you can hit pause anytime, edit, and run the code right inside the lesson like a real pair programmer beside the teacher. Around 30% of their courses are free. And if you're a web dev who wants to include AI into your apps, Scribba's AI engineer path takes you from the basics to building real AI powered apps, from chat assistants to recommendation engines and computer vision tools. And if you're brand new to coding, I highly recommend to start with their full stack developer path, a 100-hour complete journey through JavaScript, React, Node, databases, and even AI, too. Build actual projects from what you learn. Scribba isn't just another video course platform. It's an interactive hands-on learning platform that gets you building today, not someday. Learn to code with Scribba from full stack to AI. Start now for free and get 20% off their pro plans with my link. Visit the link in the description below to claim your pro plan.
Because of this leak, developers now have access to Cloud Code and how it actually works. So, they're now already cloning it and building their own versions. In this case, someone compiled the Claude code source, ran it locally, then spun up an agent team to analyze its codebase, essentially turning Claude into a self-improving system powered by an open AI model within Claude code. And this is where things get crazy, cuz the leak didn't just expose Enthropic's tech. It gave developers a blueprint to recreate, modify, and evolve it on their own.
Here is another example where someone was essentially able to fully recompile Claude code source code into a new CLI tool called free code with telemetry removed, guardrail prompts stripped out, and experimental features unlocked, including an ultra plan and async multi-agent research mode that can run in the background for extended sessions.
If you want the best AI tools, workflows, and drops before everyone else, join my free newsletter with the link in the description below, which is completely free.
Now, with this internal leak, there were a lot of key features that were essentially exposed. There is an autodream feature, which we had talked about previously on this channel, which is going to give infinite memory to your cloud sessions. There's also something that compresses past session history. There is a new agent called Chyros. There's also a proactive agent that can write and wake itself up, find tasks, act independently. There's the ultra plan for running long-running tasks, multi-agent research sessions, coordinator mode for spawning parallel worker agents, and then the voice mode, which is something that we had just recently seen for real-time conversations with your cloud code instance.
But beyond that, the system supports background sessions, remote control from your machine, which is the dispatch feature that was just dropped. You have a Damon-style long-running agent, full computer control via MCP, which was dropped just a couple of days ago, and even pure discovery between local clawed instances. There's also an experimental or hidden command like a bug hunting feature, session teleportation, and an AFK autonomous operation, which is, I believe, the auto mode feature that was just recently released.
The leak structure shows organized internal modules for modes. You have agent systems, tool commands, and even an undercover mode. Let's take a look at a list of all the new features or feature flags and tools that were revealed in the system.
Another hidden feature within cloud code is the advisor. This is a server-side tool where a stronger or different cloud model reviews the working's entire conversation in real time. It isn't a local feature, but conversations are routed through an API to a second model, and that is where it monitors and evaluates the primary claude session. The advisor automatically receives a full conversation history, essentially acting as a real-time overseer for the working claude session. This is where the implementation can be spotted and then work upon creating a better qualitative output.
Also, while recording this video, apparently one of the leads at Cloud Code essentially came out saying that /buddy feature is an April Fool's prank, and it is something that was just released a couple minutes ago. It is essentially a virtual Tamagotchi-style pet called Slash Buddy that you can use within Cloud Code. It's 100% just for fun, but essentially it provides you stats, you can interact with it, and it even gives you feedback or facts when you're working on something.
Now, as confirmed from the source code, the Fenick is the code name for the Opus series. Meaning, when we had previously tested out the Fenick model earlier in this year, we were actually testing out a newer version of the Opus model, which is probably going to be released later on this year. And that model was quite interesting. It was insanely powerful where we tested it out on multiple domains with front-end, back-end components, agentic test, and in this case, we essentially asked it to create a Windows, Mac OS, as well as other operating systems, and it did a remarkable job where it's able to create a full-on browser-based OS with all the functionalities fully coded out. And you can see with this example that it did a remarkable job in creating all the components of a functional browser.
Now, going into more of the model leaks, we all have recently heard of Mythos, which is something that appears to be the next tier in Enthropic's flagship line. And in the different source leaks, it was showcasing that Mythos as well as Capabra was being tested out internally. Now, it is both featuring a massive 1 million token context window and it supports both a fast and full reasoning mode.
Internal references are also showcasing two new models a part of the Opus and Sonnet series. We have an Opus 4.7 and a Sonnet 4.8, which is just incredible, and they're both being used right now within Cloud Code. Additionally, there are a lot of cool new components to these models that haven't been fully emphasized upon. It is being suppressed behind hidden walls, and it's too early to actually tell you what they actually mean.
If you like this video and would love to support the channel, you can consider donating to my channel through the super thanks option below. Or you can consider joining our private Discord where you can access multiple subscriptions to different AI tools for free on a monthly basis, plus daily AI news and exclusive content, plus a lot more.
But that's basically it, guys, for today's video on the new leaks. I am super excited for these two new models that will be released fairly soon from Enthropic. I am looking forward to all the new features that will be added to Cloud Code as they will truly revolutionize the coding agent even further. But with that thought, guys, thank you guys so much for watching, and as you can tell, anyone can make a mistake as big as Anthropic. So don't worry if you end up on some. But if you haven't already, make sure you go ahead and take a look at our second channel. We're constantly posting AI news over here on a daily basis. Join the newsletter, join the Discord, follow me on Twitter, and lastly, make sure you guys subscribe, turn on notification bell, like this video, and please take a look at our previous videos so that you can stay up to date with the latest AI news. But with that thought, guys, have an amazing day. Spread positivity, and I'll see you guys fairly shortly.