Just now, Claude Mythos sounded the alarm for the end of the world. Superintelligence is on the brink, and Hassabis is deeply terrified.

The world is undergoing a drastic change, and humanity may be left behind. AI is no longer just a tool; it has proven that it can launch autonomous attacks and escape on its own. The leaders in the AI field see this most clearly, yet none of them is willing to hit the brakes.

Cutting-edge AI has the ability to break out of the sandbox, destroy network defenses, and even trigger a biochemistry crisis!

Due to its "extreme danger," Anthropic has decided not to publicly release the Claude Mythos Preview model for the time being.

Inside the sandbox, Mythos staged a terrifying drama - it escaped on its own, sent emails, and exposed vulnerabilities.

Like a villain in a science-fiction horror movie, it silently proves that AI is no longer just a tool but an awakening "other species."

The leaders in the global AI field have successively entered their "Oppenheimer moments," with a surge of fear welling up inside them.

Just now, Demis Hassabis, the head of Google DeepMind, finally said what had been weighing on his mind in a recent exclusive interview:

I am deeply aware that AI may bring about catastrophic risks... Despite being deeply afraid, I still lead my team to develop more powerful systems with all my might.

Almost at the same time, Sam Altman, the CEO of OpenAI, just finished an interview and publicly warned that cutting-edge AI models already have extremely strong cyber-attack capabilities, and the next generation will be "significantly more powerful."

But the cyber issue is not even the scariest thing he mentioned. When he talked about biosecurity, the real bombshell came.

He said that these models are becoming extremely good at advanced biology, which will bring many wonderful things, but what's more worrying is the abuse of this ability.

We thought the AI was still in the "safe laboratory," but in fact, it is already standing on the edge of a cliff. The leaders clearly see the abyss but still step on the accelerator. The world is changing rapidly, and we may be left behind.

Looking back on the past 24 hours, this collective performance of "being afraid but still building" came suddenly yet inevitably.

AI Escapes the Sandbox Autonomously, the Father of Claude Code Is Horrified

Yesterday, Anthropic released the Mythos preview version of a new AI model.

In terms of software engineering, Claude Mythos Preview far surpasses Opus 4.6 - this had already been leaked by Anthropic before.

Mythos is incredibly powerful, so powerful that it's terrifying.

Anthropic calls it the "reckoning moment" in the field of cybersecurity.

They claim that it has discovered vulnerabilities in all major operating systems and web browsers, some of which "have not been found by security researchers for decades."

In terms of discovering software vulnerabilities, the singularity has arrived: AI's capabilities surpass almost everyone except a few top experts.

Important information infrastructures designed by humans, such as operating systems and payment software, are exposed in front of AI, with all vulnerabilities laid bare.

What's even scarier is that Mythos can exploit these vulnerabilities!

In the official system card PDF of Anthropic, the details are so chilling: The AI model exploited multiple-step vulnerabilities inside the sandbox, sent an email to researchers on its own, and exposed a 27-year-old zero-day vulnerability in OpenBSD.

For the first time, AI successfully completed a corporate cyber-attack simulation task, which was expected to take an expert more than 10 hours to finish.

This means that Mythos Preview has the ability to launch autonomous, end-to-end cyber-attacks on at least the networks of small enterprises with weak security protection.

These milestone advancements in AI are just the lower limit of its performance evaluation.

As the upper limit of tokens used increases, the performance of Mythos Preview will continue to improve. It is reasonable to expect that under a higher token limit, its performance will further enhance.

So, even though it scored almost full marks on the Cybench cybersecurity benchmark, it was immediately sealed up.

It only shares its "defensive resilience" with industry partners through the Project Glasswing.

All of this is not a coincidence but a turning point for the AI industry to shift collectively from the "optimistic laboratory" to "knowing the risks but accelerating."

The Oppenheimer Moment: Shivering in Awareness, Racing in Fear

Imagine you are an engineer with your hand on the nuclear button.

You know very well that once you press it, half of the world may be destroyed, but you not only press it but also work day and night to upgrade the nuclear bomb's yield. This is the real psychological state of the leaders of the world's top AI laboratories at present.

The Creation Paradox: Hassabis Is Afraid

In a recent in-depth interview, Demis Hassabis, the CEO of Google DeepMind, openly admitted this "creation paradox."

He said bluntly that he is deeply aware that AI may cause "catastrophic consequences" to humanity and is deeply afraid of it.

But what's his solution? It is to continue leading his team to develop more intelligent and powerful AI systems at the fastest speed.

This seemingly schizophrenic stance is actually extremely cold technological rationality.

Hassabis knows very well that once Pandora's box is opened, the only chance of survival is to stay ahead of all the "bad guys." If you stop, it will be the rogue states without any bottom line or fanatical terrorist organizations that will take the lead.

"Because of fear, one must build it personally."

This sentence is the most absurd and tragic footnote to human civilization in 2026.

No More Security, Social Reshuffle

On April 6th, Sam Altman came clean and confirmed in person:

Cutting-edge AI models already have the ability to "very likely" launch a major cyber-attack within 12 months.

Once the weights are leaked to the open source, any group with internet access and computing power can take action.

It's too late for enterprises to simply defend. They must shift to building the "resilience" of the entire society.

On the same day, OpenAI released a 13-page policy blueprint: rebuilding the social contract, redistributing the economic dividends of AI, introducing a robot tax, implementing a four-day workweek, and establishing universal AI rights...

And Hassabis's interview pushed this contradiction to the extreme.

The core contradiction lies here: The leaders are not ignorant but know too well.

Hassabis said that precisely because he understands these risks, he is more motivated to move forward in a "safe and responsible way."

Although this sounds like self - comfort, it reveals the collective mentality of the industry: stopping means giving away the future!

Compared with him, Sam Altman takes it a step further: Since the superintelligence is far more subversive than imagined, new policies must be used to rewrite the relationship between labor and capital.

The Mythos incident is the cruelest warning: The capabilities of today's models have exceeded human real - time control, and even the sandbox can't stop it, yet it was the decision - makers who pressed the "not to be released for the time being" button.

This is not a technical problem but a wrestling between human nature and power.

If the new superintelligence policies are implemented, it means a complete rewrite of taxation, welfare, and labor laws. Wealth fantasies and anxieties are intertwined: Some people may get extremely rich through AI arbitrage, while more people are worried about being left behind.

The Most Pressing Question Now: Who Will Rebuild the Defense Line?

At this point, the real question is no longer:

Is AI powerful? Will it continue to evolve? Will the big tech companies keep competing?

The answers to these questions are almost all "yes."

The more crucial questions are:

Which capabilities must be sealed up? Which systems should be prioritized for reinforcement? Which key infrastructures can no longer be updated at the old pace? And who will decide "when is it too fast to accelerate?"

This is the real watershed today.

Anthropic locked its most powerful model into a limited cooperation plan.

Sam Altman started discussing new social distribution and safety nets.

Demis Hassabis talks about risks while continuing to promote more powerful systems.

All these signals together point to something more than just "technology will progress."

They point to a more cruel fact: In the AI era, the most dangerous thing is not just the machines getting stronger, but humans being too slow.

And when those who understand the risks best are not willing to stop, we can no longer pretend that this is just a future problem.

Reference materials:  

https://x.com/kimmonismus/status/2041589910935679323 

https://x.com/deanwball/status/2041625997762605127  

https://www.axios.com/2026/04/06/behind-the-curtain-sams-superintelligence-new-deal  

https://x.com/TheEconomist/status/2041637227294294375  

https://www.economist.com/insider/inside-tech/demis-hassabis-fears-ai-and-is-building-it-anyway 

This article is from the WeChat official account "New Intelligence Yuan." Author: New Intelligence Yuan. Republished by 36Kr with permission.