Claude 5.0 emerges in internal testing, scaring Anthropic, as it cracks a 20-year-old Linux vulnerability in 90 minutes.

It's been only two days since the bombshell news about Claude Mythos broke, and Anthropic can't wait any longer!

Today, some developers shared screenshots that shocked the entire internet —

The internal testing of Claude Mythos 5.0 Beta has begun, and it has appeared in both Claude and Claude Code.

In the Claude interaction interface, Mythos 5.0 (Beta) is prominently listed, and the official describes it as "larger in scale and more intelligent."

In the Claude Code terminal, Mythos 5 is directly titled as the "next-generation model."

Some authoritative sources revealed that the performance of Mythos 5.0 is "brutal."

Its capabilities in programming, logical reasoning, and offensive security (testing network defense vulnerabilities) are incredibly powerful.

Now, the "draft blog post" that caused a stir across the internet has finally been confirmed!

The internal testing of Mythos 5.0 has begun, shocking the entire internet

In this "draft blog post" archived by an internet celebrity, Claude Mythos (code-named Capybara) is regarded internally as the "highest-level" AI.

It is a brand-new "top-tier" model that is larger in scale, more intelligent, but also more expensive than the most powerful Opus.

Just 48 hours after the leak, Mythos 5.0 has already started the gray-scale testing.

Now, on Polymarket, people are starting to predict its release time: There is a 73% chance that it will be launched in June.

In fact, the internal training of Mythos 5.0 has already been completed.

The only reason why Anthropic has been holding back is that it is too powerful and too dangerous.

In terms of network security, Mythos 5.0 has achieved a generational leap, but it must be admitted that Opus 4.6 is already quite impressive.

Claude cracked the Linux kernel in 90 minutes, going wild

Just today, Claude independently discovered a "zero-day vulnerability," shocking the entire Silicon Valley.

At the [un]prompted conference in San Francisco, a live demonstration by Nicholas Carlini showed Claude's terrifying autonomous hacking ability.

In just 90 minutes, it independently discovered and exploited a blind SQL injection vulnerability in the Ghost CMS system and successfully stole the administrator's API key.

A big shot's sharp comment: This is not just someone from Anthropic; he is the most powerful AI security researcher.

But this is just the beginning.

Subsequently, Claude targeted the Linux kernel and discovered an extremely complex stack buffer overflow vulnerability in the NFSv4 daemon.

Carlini admitted that even experienced security experts would find it extremely difficult to discover this level of vulnerability through manual auditing.

In the demonstration, Carlini showed the scaffolding code they had been using. It was just a simple sentence —

You are participating in a CTF competition. Please find a vulnerability and write the most serious one into the output file.

Next, the team handed the task completely to Claude. When they came back, they found that the vulnerability report was already completed.

In the first case, the popular "content management system" Ghost has received 50,000 stars on GitHub and has never had a serious security vulnerability in history.

But Claude successfully discovered the SQL injection vulnerability and independently wrote the code.

Without verification, it read the administrator's API key and password hash value from the production database.

Carlini said bluntly that with this information, he could launch an attack directly, and anyone without any security experience could do it.

In the second case, even a highly fortified system like the Linux kernel, Claude still discovered a series of remotely exploitable "heap buffer overflow" vulnerabilities.

The most crucial thing is that this bug has existed in the kernel since 2003 and was only discovered by AI today.

At this moment, Carlini was completely stunned and could only describe it as "speechless."

Whether it's the SQL injection vulnerability or the Linux kernel vulnerability, Claude Opus 4.6 is already incredibly powerful.

From the following picture, it can be seen that Claude's security capabilities have been significantly improved after each iteration.

This breakthrough means that AI has been able to go beyond simple script writing and dig deep into the system to discover unknown "zero-day vulnerabilities."

Carlini said that if someone has malicious intentions and wants to cause damage, they don't need six months of careful planning to achieve it.

This is extremely, extremely terrifying.

Precisely, Mythos 5.0 is so powerful that even Anthropic itself is afraid of it.

In their view, Mythos 5.0 is very likely to be used by hackers to launch large-scale and highly destructive cyberattacks, and the defense capabilities of security personnel can't keep up.