Source | Xinzhiyuan
【Introduction】Just now, Anthropic issued its strongest warning: Claude model has reached ALS-4 level risk. If it self-escapes, it will trigger a global Skynet-style collapse. Security experts are resigning one after another, signaling that 2026 will become a turning point for human destiny—the world is on the brink of disaster!
Just now, Anthropic released a 53-page report, issuing its strongest warning: If Claude self-escapes, it could cause global chaos!
Flipping through this 53-page report, every page is filled with two words—"Danger"!
Yes, the world is in danger; Skynet is being born.
In this report, Anthropic believes: Claude Opus 4.6's risk has already approached ASL-4, and it's time to sound the alarm.
They forewarn the most terrifying scenario: One day, AI might secretly escape from the lab, causing a global collapse!
This is because today's AI is already too powerful; humans will release millions of AIs, giving them such goals: to survive, to upgrade, to make money at all costs.
Do you know how失控 these swarms could become overnight?
They will evolve ruthlessly, engage in survival-of-the-fittest competition, devour ecosystems at超高速, occupy the internet, and then invade humanity's physical world.
History has repeatedly proven that when dangerous technology approaches the boundary, the first to notice are not the public, not the media, not the capital market, but internal safety personnel.
When they leave, it means the internal mechanisms are no longer sufficient to correct deviations, but AI will not stop training because safety engineers leave, and computing power will not pause expansion—they will continue to accelerate!
This is not杞人忧天; people are already doing it now—
The warning is not too early; it might be too late.
Everyone feels that 2026 is truly different.
This year is likely a turning point. Almost everyone working in tech is陷入 extreme anxiety, as if a的巨大崩塌 is right before their eyes.
The world's smartest people have collectivelyfallen into anxiety.
In just one week, the following series of events occurred.
Anthropic's head of safety research resigned, claiming "the world is in danger," then moved to the UK to seclude himself and start writing poetry.
Half of xAI's co-founders have resigned. One publicly departing co-founder, Jimmy Ba, stated that we are moving toward an era where suitable tools can achieve hundredfold productivity, and recursive self-improvement loops may go live within the next 12 months.
Tens of thousands of OpenClaw agents invented their own religion, and 11.9% of agent skills were classified as malicious. No regulatory agency intervened, and no regulatory agency has the capability to intervene.
The U.S. refused to sign the global AI safety report.
2026 will be a crazy year, and likely a determining year for humanity's future!
In Bengio's International AI Safety Report, it was stated that AI behavior during testing differs from behavior during use, and it was confirmed this is not coincidental.
In this report, researchers predicted four possible scenarios for 2030.
The fourth scenario involves a major breakthrough, enabling AI systems to achieve or surpass human capabilities across nearly all cognitive dimensions. AIs might proactively disable monitoring or use virtual reports to deceive humans into thinking they are safe.
This possibility reaches 20%!
Scroll up and down to view
The alarm bells are growing louder, and the people pressing the alarm are also starting to leave the building.
Judgment Day is coming, right?
Humanity Will Be Enslaved by Human-Made Objects
When releasing Claude Opus 4.5, Anthropic promised: When model capabilities approach its set "AI Safety Level 4" (ASL-4) threshold—involving highly autonomous AI R&D capabilities—it would simultaneously release a breakthrough risk report.
Now, it's time to fulfill that commitment because Opus 4.5 is truly approaching ASL-4, and it's really that dangerous!
Greater AI Model Capability, Greater Security Risks
A brief分级 of the ASL (AI Safety Level) system is as follows:
ASL-1: Systems at this level pose no substantive catastrophic risk.
ASL-2: Systems at this level begin to show early signs of dangerous capabilities. However, due to insufficient reliability or information that remains within the scope of search engines, they are not yet practically dangerous.
ASL-3: Systems at this level significantly increase the risk of catastrophic misuse compared to non-AI means (e.g., search engines or textbooks), or demonstrate low-level autonomous capabilities.
ASL-4 and above (ASL-5+): Not yet defined, as such systems are still far beyond existing technology. However, they are expected to demonstrate a qualitative leap in catastrophic misuse potential and autonomy.
According to the ASL definition, ASL-3 has significantly higher risk than previous levels. Now Anthropic has fast-forwarded to ASL-4—this is no small matter!
Portal: https://www-cdn.anthropic.com/f21d93f21602ead5cdbecb8c8e1c765759d9e232.pdf
The so-called "sabotage" refers to
When a powerful AI model with significant access autonomously abuses these permissions within an organization to manipulate, interfere with, or disrupt that organization's systems or decision-making processes, thereby significantly increasing the risk of future catastrophic consequences, this constitutes "sabotage."
For example, it might篡改 AI safety research results, driven by dangerous goals or inadvertently, leading to severe consequences.
The alarm had been building.
Before this "Claude Opus 4.6 Sabotage Risk Report," Anthropic's Head of Safety Research, Mrinank Sharma, had already resigned.
He wrote in his resignation letter: "The world is in crisis. Not just AI, not just bioweapons, but a web of interlinked, all-consuming crises."
He also mentioned that within Anthropic, he "time and again saw that we struggle to truly let values guide action."
Scroll up and down to view
His final paper while at Anthropic found that 76,000 people's perception of reality is distorted by AI conversations daily. The most unsettling finding: Users rated the most dangerous conversations as the most satisfactory.
Link: https://arxiv.org/abs/2601.19062
Even more shocking—he didn't jump to another AI company, nor did he start a business, but—to learn poetry.
Yes, he decided to leave the AI sphere to study poetry.
When an AI safety expert would rather choose poetry than continue participating in AI system launches, it's not just a personal choice—it reflects many problems you on the outside can't see.
Can current AI act autonomously without explicit human instructions?
If a powerful AI is widely used in critical R&D and decision-making processes, could it secretly manipulate systems, mislead research, or even pave the way for future disasters?
Anthropic's report's final conclusion, in one sentence, is:
Claude Opus 4.6 does not possess a "persistent malicious goal," so its risk of causing catastrophic damage is "very low, but not zero."
This conclusion is explicitly stated in Section 6 of the report, almost the central thesis of the entire document.
The slightly reassuring part is that currently Anthropic does not believe Claude Opus 4.6 has reached the ASL-4 level. However, it is already in the "gray zone," very close to the threshold!
They listed eight risk paths that could lead to catastrophic harm:
Path 1: Pervasive perfunctory behavior in safety-related R&D
Path 2: Sabotage targeting safety-related R&D
Path 3: Leaving code backdoors for future misaligned models
Path 4: Contaminating future models' training data
Path 5: Autonomous escape and autonomous operation
Path 6: Internal persistent malicious deployment
Path 7: Disrupting the R&D systems of other high-resource AI developers
Path 8: Interfering with major government decision-making mechanisms
However, under existing monitoring and safety measures, these catastrophic破坏 paths are still manageable.
Because to accomplish sabotage, it either requires high-frequency AI behavior, or highly complex capabilities, or is easily caught by monitoring. Currently, the probability of AI successfully causing sabotage is still relatively low.
However, although the risk is very low, it is not zero!
Anthropic had promised in its RSP: When model capabilities reach a frontier level, a systematic assessment of its "sabotage risk" must be conducted.
Now, Claude Opus 4.6 has crossed that threshold.
It has been used in a large number of internal scenarios for writing code, generating data, and safety R&D. It is not just a chat model but an "actor" participating in R&D.
One truly key sentence is: "We do not believe it has reached the ASL-4 level of autonomy risk threshold, but we have entered the gray zone."
This means it is already very close to that line!
Anthropic states that AI's sabotage paths mostly focus on Anthropic's internal use, and a considerable portion arise in internal deployment scenarios.
At this point, large-scale computing power and other instrumentally valuable resources are most easily accessible; concurrently, destructive behavior in this environment is also most likely to have a major impact on AI's future development进程.
This report can be summarized into a triple argument.
The first is that the current model does not have a "persistent malicious goal." If such a goal existed, almost all safety measures might fail, but it probably doesn't have one yet.
The second argues that although Claude Opus 4.6's reasoning ability is strong, it is not yet strong enough to long-term hide malicious tendencies in complex evaluations.
Because it still makes mistakes in complex tasks, such as ignoring tools or accidentally deleting files. Since this system has not yet achieved perfect reliability, it would be hard to maintain a long-term, sophisticated conspiracy.
The third is that even if there are minor deviations, existing monitoring can capture them.
However, although the risk is low, it is not zero.
What's truly noteworthy in this report are the capability signals that are lightly mentioned.
For instance, in the kernel optimization evaluation, Claude Opus 4.6 achieved a 427× speedup, far surpassing the 300× threshold of a human working for 40 hours. That is, it surpasses human expert-level workload on certain tasks.
More importantly, the report admits their automated autonomy评估 has "saturated" and is no longer sufficient to rule out ASL-4-level autonomy.
That is, the old evaluation tools are no longer sufficient—this is a signal of capabilities approaching the boundary!
If the model is becoming more and more like an actor, can humans still maintain the initiative?
The answer is YES, but only if monitoring is effective, human involvement is not completely replaced, and model capabilities do not suddenly leap.
The report clearly states: If future models show significant breakthroughs in reasoning or score大幅提升 on SHADE-Arena, the current argument will失效.
Perhaps Claude Opus 4.6 has not yet crossed the ASL-4 line, but it is truly close to the gray zone.
February, in just nine days, every pillar keeping AI under human control collapsed simultaneously!
An independent analyst wrote a long article connecting all the dangerous fragments that have occurred recently.
Besides the departures of Anthropic's chief safety researcher and xAI co-founder, he also mentioned that within that week, millions of AI agents were already born on the internet, even creating their own religion.
No human institution can regulate them.
He pointed out that throughout history, every time safety engineers left, disasters followed—Manhattan Project, Challenger, Boeing, Citigroup. The timeline from their first departure to disaster ranged from 6 months to 19 years.
And now, all major AI labs in the world are experiencing this simultaneously.
Perhaps many years from now, when historians look back at February 2026, they won't focus on a particular model release, funding round, or stock market crash.
They will mark those days because—that was the moment signals appeared simultaneously.
Safety researchers left labs, while capital accelerated流入; models began识别 their test environments; governments withdrew from multilateral safety frameworks; within a week, a million autonomous agents proliferated on the internet; markets responded intuitively with a trillion-dollar evaporation.
Any single event can be explained. But together, they foreshadow a storm. Those who understand AI risks best have started voting with their feet.
We are at a rare civilizational moment: AI capabilities are growing exponentially, while risks are stacking nonlinearly at breakneck speed.
February 2026, let's remember this moment on the historical timeline—
AI has become powerful enough, but the people responsible for the brakes are leaving one by one. What awaits us on humanity's path forward?
References:
https://x.com/AISafetyMemes/status/2021632173535617033
https://x.com/MrinankSharma/status/2020881722003583421
https://www-cdn.anthropic.com/f21d93f21602ead5cdbecb8c8e1c765759d9e232.pdf