Claude fixed bugs by itself overnight with a major update, and its PR status turned green. Programmers around the world: "Slacking off" is now officially legal.

[Introduction] Claude has been updated again! Early this morning, Claude Code launched the "Auto-fix in the Cloud" feature: it can fix bugs autonomously and keep PRs always green. The day when programmers can finally free their hands has come.

Claude's frequency of being updated daily has shocked the entire network!

73 products in 52 days, unprecedented

Well, this morning, Claude Code finally launched the "Auto-fix in the Cloud" (auto-fix-in the cloud) feature.

Whether using the web version or the mobile version, Claude can automatically follow up on PRs.

Once the CI reports an error or someone proposes modification suggestions, it will independently troubleshoot problems, write code, and push it to the repository, ensuring that the PR is always "green".

This means that programmers don't need to be online all the time. By the time they finish a cup of coffee and come back, the work is already done!

The "Auto-fix in the Cloud" feature is in line with the "Auto Mode" (auto-mode) launched a few days ago, completely freeing developers' hands.

Some netizens said that the CI changed from red to green without any human intervention throughout the process. This is either the future or the beginning of a horror movie.

Even more amazing is that a doctor from the ELLIS Institute used Claude Code for automated scientific research and created an autonomously operating scientific research closed-loop system - Claudini.

After 56 iterations, a new jailbreaking algorithm that can defeat more than 30 types of GCG attacks was successfully discovered.

Claude Code can fix bugs and submit code autonomously, conduct scientific research and iterate and evolve autonomously. A real ASI closed-loop has been completed.

Claude Code Fixes Bugs Autonomously, and Programmers "Slack off" Collectively

For developers, the most headache moment is when they find the CI turns red after submitting a PR.

Even for minor syntax errors or simple annotation suggestions, they often need to reopen the computer, switch branches, make modifications, and submit again.

However, the "Auto-fix in the Cloud" feature launched by Claude Code today will completely change the game rules.

Undoubtedly, the core of this feature lies in "fully automated" and "remote".

In the past, AI-assisted programming often relied on the local environment. Now, Claude Code can directly take over GitHub PRs in the cloud.

Once the auto-fix is enabled, Claude will monitor the status of the PR in real-time:

• Automatically investigate the cause of failure: When the CI test fails, it will actively analyze the error log.
• Respond to review comments: For the comment suggestions in the PR, it will understand the requirements and propose countermeasures.
• Automatically push the fix: Directly submit the fixed code with detailed explanations to ensure that the PR status always remains "green".

How to get started quickly? The official provides three ways -

For the PRs created by the web version of Claude Code, directly click CI > auto-fix; on the mobile version, directly tell Claude "auto-fix" on the phone.

In addition, you can also throw any PR you want to monitor to Claude, and then the auto-fix can start.

CC engineer Lydia Hallie strongly recommends developers to enable "auto-fix" and go slack off!

For most developers, coming back to a PR that is ready to be merged after leaving for a while is simply a dream.

Now, a large number of programmers are celebrating online. The day when their hands are truly freed has come!

56 Major Iterations, Self-evolution in Scientific Research

Claude Code not only competes in coding but also in scientific research.

Also today, research teams from institutions such as the ELLIS Institute Tübingen released a shocking progress -

They deployed Claude Code in an automated research (Autoresearch) cycle, allowing the AI to independently search for new LLM jailbreaking algorithms.

Paper link: https://arxiv.org/pdf/2603.24511

The results prove that this automated pipeline called Claudini not only works but also has excellent results.

In the attack test of the "harmful queries" of GPT-OSS-Safeguard, after about 56 iterations, Claudini independently developed a new algorithm that can defeat all benchmark tests.

Moreover, the attack success rate (ASR) on the validation set is as high as 40%, while the highest among all previous benchmark models was only 10%.

Even more terrifying is that in the test of the Meta-SecAlign-70B model that has undergone "adversarial training", the 𝚌𝚕𝚊𝚞𝚍𝚎_𝚟𝟼𝟹 method discovered by Claudini achieved a 100% attack success rate.

In this process, Claude Code behaves like an extremely experienced scientific research hacker.

It is extremely good at recombining existing attack strategies, radically adjusting hyperparameters, and squeezing out the last bit of performance through "reward hacking" when the progress reaches saturation.

This research sends a clear signal that incremental security research can now be completely "automated".

When Claude starts searching for defense vulnerabilities 24/7, the AI has learned to make its own "digital keys" to unlock its own locks.

Recursive Self-improvement, 100% of Code Written by AI

The Claude Code team updates almost daily and releases products crazily.

Starting as a side project, Claude Code was officially launched in February 2025. Six months later, its annualized revenue was approaching $1 billion.

Now, its daily active users have doubled, and it accounts for 4% of the public submissions on GitHub.

Behind this, it's not the result of the "genius" development team staying up all day. Instead, Claude Code has truly achieved self-evolution.

Boris Cherny, the father of CC, has publicly stated more than once that 100% of his code is written by Claude Code.