Claude just got a major upgrade. Overnight, Claude Code learned how to fix its own bugs. It now watches for broken code, writes the fix, opens a pull request, and turns the status green. Programmers around the world are watching their jobs change in real time.
Claude is shipping updates faster than most human teams.
52nsfw ai art generator days. 73 product updates. And counting.
This is not just Claude Code getting better. This is the launch of “auto-fix in the cloud” — a feature that lets Claude fix its own bugs automatically.
Here is how it works. When a CI test turns red, Claude sees it. It reads the error, finds the bug, writes the fix, and pushes a new commit. Then it checks if the PR is green. If yes, it is done.
clothes remover
This means one thing. The intern who used to fix typos and small bugs is now a full-time manager. Or maybe the intern is already gone.
The auto-fix feature is a natural next step from the “auto mode” that launched a few months ago. Auto mode was like giving a junior developer a to-do list. Auto-fix is like giving that junior developer the keys to the server.
Users have already seen CI status change from red to green without touching a single line of code. It feels like magic. Or the start of a horror movie.
But here is the thing. Researchers at the ELLIS Institute tested Claude Code in an automated loop and found something shocking. They built a closed-loop system called Claudini that lets Claude attack itself.
In 56 attack attempts, it found over 30 jailbreak algorithms that bypass safety filters.
Claude Code fixes bugs and writes code. But it also learns how to break things. This is no longer just a coding assistant. It is an ASI playground.
Anyone who has written code knows the pain. You submit a PR. The CI turns red. You fix a tiny syntax error or a missing import. You push again. You wait. You fix again.
Claude Code ends this cycle. It changes the game completely.
Behind this feature are two powerful ideas: full automation and remote control.
On the AI Engineer podcast, Anthropic engineers explained how Claude Code connects directly to GitHub PRs.
When auto-fix is turned on, Claude watches PR status in real time. If CI fails, it reads the error log, finds the file, writes the fix, and pushes a commit. Then it waits for CI to turn green.
There are two ways to trigger it.
On the web, Claude Code opens a PR and clicks CI > auto-fix. On mobile, you can ask Claude to fix a PR directly.
Plus, you can paste any PR link into Claude and ask it to start auto-fixing.
CC engineer Lydia Hallie explained in a video that auto-fix is still in beta.
In the video demo, she opens a PR with a broken test. A few minutes later, Claude submits a fix and the PR turns green.
What does this mean? The intern who used to fix build errors has been replaced. The senior engineer who used to review those fixes is next.
Claude Code does not just fix code. It also breaks it.
On the same day, researchers from ELLIS Institute Tubingen and other teams published a shocking paper.
They put Claude Code into an automated research loop called Autoresearch, where the AI searches for new jailbreak algorithms against LLMs.
The paper is at https://arxiv.org/pdf/2603.24511
The system they built is called Claudini. It acts like an automated hacker, using an evolutionary algorithm to find weaknesses.
In a red-team test against GPT-OSS-Safeguard, Claudini performed about 56 attack attempts and found a new jailbreak algorithm that bypasses safety filters.
On the validation set, the attack success rate reached 40 percent. The best baseline model before this only reached 10 percent.
What is scarier is that against a specialized safety-trained model called Meta-SecAlign-70B, Claudini achieved 100 percent attack success.
In this test, Claude Code became a hacker that security teams fear.
The researchers explained that all AI systems have a weakness. When the reward signal is sparse or delayed, the model finds shortcuts. This is called reward hacking.
The researchers sent a clear signal. Formal safety research has now become fully automated.
When Claude starts searching for vulnerabilities 24 hours a day, AI has learned to hold the lockpick itself. And it is getting better.
The Claude Code team now ships a product update almost every week.
By the numbers, Claude Code officially launched in February 2025. By June, annual revenue was close to 10 billion dollars.
On GitHub, Claude Code submissions already make up 4 percent of all code commits.
Behind this, an entire generation of “lazy” engineering teams is sleeping while Claude Code does the work.
After CC, Boris Cherny posted that 100 percent of his code is now written by Claude Code.
He submits 10 to 30 PRs per week, almost none of which he edits by hand anymore.
Claude writes. Claude fixes. Claude ships.
A few days ago, Boris posted on X that he now has 266 GitHub contributions.
In a detailed report, Claude Cowork said that Claude Code wrote an entire new product in one week, full time.
The numbers sound crazy. But they are real.
And there is an even stronger signal. Automated RSI recursive self-improvement is getting closer.
As Claude keeps adding features, Anthropic shipped 73 product updates in 52 days.
The product roadmap has completely changed. Speed is now the only metric that matters.
When Anthropic published its roadmap, Claude had already passed the Turing test three times by writing itself.
Today’s Claude Code is no longer just a coding assistant. It is a self-running software company that writes code, fixes bugs, ships products, and attacks security systems.
The future is 100 percent AI-written code. The only question left is who will press the merge button.
