Late last night, Anthropic’s top-secret Mythos benchmark scores leaked online. The numbers were so strong they broke records on the spot. But the leak also revealed something much stranger. Hidden inside the leaked Claude Code source code were details about a model called Capybara. And that was not all. Someone found proof that Anthropic had been slipping secret instructions into the code.
In the past 24 hours, the AI world has been caught between shock and laughter.
It all started with Claude Code.
Earlier, someone leaked the source code of Anthropic’s terminal tool on GitHub. Developers rushed to fork the repo. They dug through Python and Rust files. What they found was a gold mine.
At first, everyone thought this was just an internal test that had gone public by accident. A simple mistake.
No one expected the second leak to be even bigger.
Just now, Anthropic’s flagship model Mythos had its benchmark scores leaked. The numbers were shocking.
Unlike the Claude 4.x/5 series, Mythos is positioned as a true flagship. It aims for the highest possible performance. Leaked data shows it is likely the first model from Anthropic to reach the top tier.
The leaked info compares Mythos against the current strongest model, Opus 4.6. On every key metric, Mythos shows clear gains.
Some of the numbers are so high that Mythos looks almost unbeatable. On some tasks, it scores in the top tier with almost no room left to improve.
On Terminal-Bench it hits 78.4%. On SWE-bench it reaches 87.4%. These numbers prove Mythos is built for serious work.
But the funny part is what the leak also revealed.
Someone found that Anthropic is using a hidden AI watermark called synthid. This is a secret signal that tells whether content was made by AI.
The watermark release also included accuracy checks.
After the first leak, some people guessed Anthropic might cancel Mythos to stop the buzz.
Source Code Leak Reveals Capybara Model Details
But there is more. Inside the leaked Claude Code source code, people found details about a model called Capybara that Anthropic had never talked about.
These clues hidden deep in the code point to a model that is still in early testing. The code comments suggest Anthropic is building a new model family that sits between the current lineup and the top tier. Some people think this is the early version of a model that will bridge the gap.
celebrity ai nudes
At the same time, people also guess that Mythos is the real test version. Some leaked code even points to a model name called capabara-v2-fast with support for 1 million tokens of context.
The 1M context window with auto-optimization means this model can read an entire porn ai generator book in one go. That is a game changer.
Although it carries the “fast” label, Anthropic’s internal naming style suggests this is likely a strong flagship model.
Hidden Prompts: A Machine Full of Secrets
The most interesting part is not the model names. It is what Anthropic hides inside the code to control the model.
Researchers found that Prompt Shape, Tool Use, and turn boundaries are all packed with hidden rules. Inside the Capybara code, there is a special stop signal that tells the model when to stop talking.
In other words, the model is not fully grown yet. It is still being shaped by human hands.
To fix a known bug, Anthropic did not simply retrain the model. Instead, they built a layer called Prompt Surgery on top of it.
First, they added hard safety borders. These are like guard rails that keep the model from going off track.
Then, they used special blocks called Sibling Blocks to position related content.
On top of that, they added Reminder Text. This text is injected directly at the boundary to make sure the model remembers the rules.
They even added empty space tricks. By filling blank areas with non-empty markers, they stop the model from getting confused in empty spaces.
All of this adds up to a machine that is more puppet than free mind.
Internal Kill Switch: The Tengu System
There is more. Anthropic has an internal system they call the Gray Switch.
This means that behind Capybara’s friendly face, there is a hidden control layer.
It works like a kill switch. If something goes wrong during rollout, they can pull the plug and roll back instantly.
The leaked code also includes A/B test data.
What is funny is that both users and Anthropic staff are treated as test subjects. New features only go to outside users after passing internal tests.
Leaked Code Shows Anthropic Slipping Secret Commands
So far, the whole world has seen the leaked source code. But some people found something darker. Anthropic is not just building models. They are slipping secret instructions into the code.
In the model world, the common view at the bottom layer is that data is the only thing that matters. You grab a lot of data, train the model, and then you are done. But this leak shows that Anthropic thinks differently.
This is not just about the leaked Claude Code source code. Anthropic has been fighting to keep their secret sauce hidden. They sent DMCA takedown notices. They tried to bury the leaks.
But the leaks keep coming. And each time, we learn more about how the sausage is made.
At the same time, Claude is also learning to hide its own tracks. Watch closely and you will see a set of hidden tool instructions.
These instructions tell the model not just how to answer questions, but how to grab data from the backend. Some people call this poison in the training data. Others call it a hidden backdoor.
Either way, users are training the model without knowing what they are really teaching it.
To stop people from talking about it, Anthropic packed the tool use logic with so much detail that no one could read it easily.
But the leaked code shows the full picture. Every tool call, every hidden rule, every secret path is now out in the open.
From these clues, we can also guess what Anthropic is really testing. They are not just chasing better scores. They are building a game of hide and seek. In complex tool use, the model must learn to act like a formal agent without becoming one.
Although no official SKU has been released yet, the Capybara name is already showing up in the wild.
So the question is: Is Capybara the bridge between Claude 3.5 and the new 4.0 series? Or is it something completely different?
Anthropic’s Strange Silence
The funny thing is that after each leak, Anthropic stays very quiet. They do not post on social media. They do not hold press conferences. They just quietly send DMCA takedown notices to GitHub and ask for thousands of repos to be removed.
But the media got hold of the story. Anthropic’s response to the leak was to treat it as a new form of security risk. They called it a safety leak.
After Claude Code leaked, Boris Cherny also posted about a bun issue. He wrote just one line. “It was just a coding mistake.”
But people inside Anthropic say the real rule is simple. If you leak, you are out. No second chances.
The AI community has reached a strange conclusion from all these leaks. Claude Code may have been opened by accident, but it was actually a gift. The leaked source code is now being studied by people who were never meant to see it.
Before the leak, the open source community was already building their own versions. Now, with the real details in hand, they are building even faster.
Why do Anthropic’s products always look so clean on the outside but so messy on the inside? The answer is simple. They are built with Python and TypeScript, but the real work is in the prompt engineering.
Prompt fine-tuning is the one thing you cannot copy. Model weights can be stolen. But these tiny prompt details are the real secret.
Some developers joke that no matter how good the open source system looks, you cannot copy the soul. The Cursor model has already proven that even with someone else’s model, as long as you have good product design and prompt engineering, you can still create a killer product that no one can leave.
So the Claude Code leak is actually a lesson. It shows us that in the AI world, open source is not just about code. It is about trust. It is about who gets to hold the keys.
In the future, whoever controls the open source foundation will control the next wave of AI products. And right now, that race is wide open.
What Comes Next
Anthropic is walking a tightrope.
On one side, they sell themselves as the safest AI company. A new kind of Google. They publish research on AI safety. They hire top safety researchers.
On the other side, they are racing to build the strongest AI as fast as possible.
So far, they have managed to keep both stories going at the same time.
But with each leak, the mask slips a little more. The public is starting to see that the safest AI company is also the most secretive.
Anthropic has already leaked over 3000 internal files. Some of them include draft blog posts.
After all this, Anthropic did admit one thing.
The model family called Capybara is real. It is still in testing. It is a bridge between research and safety.
But a small group of people focused on AI safety and open standards are asking a bigger question. Who gets to control the keys?
The truth is, Anthropic did not stop the leaks.
And while the leaks keep coming, Anthropic’s model empire is already starting to shake. One leak at a time, the public is learning how the magic trick works. And once you see the strings, you cannot unsee them.
For a company that sells trust, that might be the most dangerous leak of all.
