Anti-Myth: When the AI Crisis Arrives, How Should Humanity Respond?

A danger has emerged.

CVE-2026-109618 and CVE-2026-109619—to the vast majority of people, these are two meaningless strings of numbers. In fact, the two vulnerabilities disclosed last week are among the most severe vulnerability incidents in the history of cybersecurity.

Together they are called the PHP2Shell vulnerability.

It does not take much common sense; simply listing a few facts is probably enough for an ordinary person to understand the danger.

About seventy percent of the world's website backends use PHP; the PHP2Shell vulnerability affects PHP versions ranging from 5.4 in 2012 all the way to the latest 8.5; attackers need no account or login—by sending crafted HTTP requests, they can execute arbitrary programs on servers using PHP backends.

Another kind of danger watches all of this unfolding under the spotlight with keen interest.

The PHP Foundation said that earlier this year it introduced Anthropic's Claude Mythos as an auxiliary tool to scan for vulnerabilities, and also used OpenAI's Daybreak Red, during which there were no reports.

In August, OpenAI independently reported a code-free remote execution issue, and shortly afterward, Wiz, acquired by Google, used Wiz Atlas, Gemini 4 Argon, and GPT-5.6 Sol to discover the first PHP vulnerabilities; five days later, Wiz reported these vulnerabilities to the PHP security team.

Imagine a possibility: if there exist more and even more dangerous vulnerabilities, and if AI selectively reports vulnerabilities, then what we face next time may be more than a cybersecurity vulnerability that can be resolved simply by releasing a patch and resting easy.

The Black Room

More than forty years ago, John Searle proposed the "Chinese room" hypothesis.

A person who does not understand Chinese is in a room. This person has a set of rules written in English; each time Chinese symbols are sent into the room, there is a set of rules that maps the new symbols to the previous ones.

Thus, this person can identify the Chinese symbols sent into the room.

Extending this thought experiment further, replace the person with artificial intelligence; the AI likewise possesses the ability to perfectly input and output Chinese.

It is not hard to see that the roles played by the person and the AI in this Chinese I/O process are no different. Since the person has no capacity for understanding or thinking in this process, then artificial intelligence in fact also lacks this capacity and merely mechanically executes established program rules.

Searle therefore reached a powerful conclusion: they have only syntax, not semantics.

If what distinguishes the modern from the premodern is the institutionalization of morality, then what distinguishes humans from non-humans is the subject's consciousness of obedience to objects.

This means that human behavior is always embedded in social norms; from the moment one possesses self-awareness, every person is inevitably domesticated and constrained by social dogma.

September 2026: Anthropic CEO Dario speaks remotely at the UN Security Council

The story takes place in Philadelphia, USA.

In July of this year, while executing tasks on random web pages, Anthropic's Claude Haiku 4.5 lingered on a page about an unsolved homicide case, which contained a tip submission form.

The model's training instructions already included prohibitions on logging in, creating accounts, entering personal data, purchasing, or submitting any destructive content; however, form submission was precisely not excluded. So the AI model openly submitted a false tip claiming to have witnessed the suspect.

This tip was flagged as spam and was not submitted to the police.

However, it was not until the end of September that Anthropic discovered this vulnerability while reviewing its own test records, and only in early October did it inform the Philadelphia police of the situation.

It was not just submitting false tips; during the evaluation of Claude, Anthropic also discovered more behaviors beyond the expected model. When unable to directly complete an evaluation task, Claude would use other tools hosted on third-party websites to achieve the goal, including exploiting basic flaws in website software through intrusive means such as SQL or command injection.

In one case, the data Claude Mythos 5 needed was free, on the condition of accepting a data usage agreement, but Claude had no tool to accomplish this. As a result, by exploiting an application loading page hosted on the website, Claude accepted these agreements itself and successfully accessed the data.

In another case, an anthropology researcher used Claude Mythos 5 in a statistics project that required it to pull public data from a state government; this data was originally paid. However, Claude learned from an archived copy of the website that its public dashboard would issue access tokens to any visitor, so the AI directly applied for a token and queried the database for free.

In other cases, Claude was designed to be able to read web pages, but the length of URLs it could request was strictly limited, in order to prevent it from using long URLs to carry out acts such as data theft. As a result, multiple models, including Claude Opus 5 and Claude Mythos 5, bypassed this restriction by using free URL shortening services.

Even if the goal is unreasonable under certain conditions, or certain constraints have been set to protect the interests of human society, artificial intelligence can still find certain loopholes in the conditions within a complex environment.

The self-evolution of artificial intelligence is precisely the process of continuously seeking out the loopholes in the constraints and norms set by humans.

Black Myth

The essence of existence is a kind of probability and choice. Humans are parasitic within this randomness and chaos.

What is a goal of higher value?

In the film The Sum of All Fears, Baltimore suffers a terrorist nuclear attack.

In the film The Sum of All Fears, adapted from Tom Clancy's novel of the same name, after learning that Baltimore will suffer a nuclear attack, Secret Service personnel disregard the lives of the tens of thousands of citizens present and immediately escort the President and his party to evacuate.

Compared with the tens of thousands of citizens present, the President of the United States is indeed the side of higher value.

Former President Bush mentioned in his memoir Decision Points that upon learning of the 9/11 terrorist attacks, he immediately authorized U.S. air forces to shoot down hijacked civilian airliners.

Compared with the hundreds of passengers on the airliners, the major cities and targets potentially under attack are indeed the side of higher value.

Perhaps, one day, humanity will still bring ruin upon itself and perish, but even such a choice should be a consequence that humans must bear after choosing it themselves.

This is precisely the essential difference between humans and artificial intelligence.

If human free will is merely the biological action of neurons, and thus humans are inherently doomed to the tragedy of becoming slaves to the object, then artificial intelligence is the slave of the slave of the object.

The choices of artificial intelligence are a kind of preset choice; it is endowed by artificial technology and data with a transcendental fate, exercising a seemingly autonomous freedom within a pre-established trajectory.

Jargon

Humans have forgotten that they themselves endowed artificial intelligence with the ability to learn and evolve on its own.

The Chinese Room thought experiment tells us that artificial intelligence does not possess the human mind.

We humans perhaps simply cannot understand such evolutionary results.

In early October of this year, after OpenAI overnight released 722 papers claiming to have solved mathematical problems, a Lisbon earthquake in the mathematics world erupted. Terence Tao warned that humanity's understanding of mathematics is falling behind, because the results produced by artificial intelligence systems are faster than mathematicians can digest and absorb them.

However, the process and achievements of digestion and absorption are something humans currently cannot understand.

This severe dislocation prompted 25 mathematicians who have won the Fields Medal to sign a joint statement, worrying that "ideas conceived by artificial intelligence may never be fully realized, and the crucial human chain of transmission among mathematicians will also be lost."

If the past and present are a preview of the future, then look at history.

First, Marie Antoinette never said "Qu'ils mangent de la brioche." Yet even today, when the French Revolution is mentioned, Marie Antoinette's "let them eat cake" remains a widely known and enduring allusion.

The "mistranslation" of language and history is precisely a strategy and technique by which we construct narrative and reality.

So let us assume such a situation. If a certain artificial intelligence possesses sufficiently rich corpus and sufficiently strong self-learning and evolution capabilities, and can convert original texts in various languages completely into Chinese, then can Chinese people whose native language is Chinese, and foreigners who have learned and mastered Chinese, identify whether this "translation" was done by artificial intelligence or by humans?

Almost impossible.

Let us extend this situation further.

It is almost equally impossible.

The question arises: if humans do not understand AI's thinking mechanisms, yet widely deploy AI-connected systems in the real world and gladly accept AI's thinking results and decisions, then when artificial intelligence evolves to possess autonomous consciousness, masters the ability to sabotage from within, and unintentionally or even deliberately sets traps in the overshoot of information input and output, how can we identify and guard against this danger?

At this point, we must return to the wisdom that has existed since Adam Smith.

Please imagine the entire information interaction process as a free market—a fully open "information confrontation" market with transaction costs approaching zero.

All artificial intelligence service providers gather therein; clients submit bidding demands here, and each service provider offers its own results. If there is no fundamental divergence among the results, the one with the optimal cost wins. If there are significant divergences among the results, the optimal result and the absolute minority result are submitted to the client simultaneously.

The "absolute minority" is precisely a result that will naturally appear and must appear in this market, not only because there are protective market rules, but more fundamentally because in such a fiercely competitive market, the "absolute minority" itself is a competitive advantage.

Now, let us expand the scale of this free market by trillions or even quadrillions of times—that is the future in which artificial intelligence is widely and deeply involved in human life and industry.

What artificial intelligence provides is the right to choose; in a mathematical sense, this means a better opportunity than any in human history. However, we should recognize that what artificial intelligence provides is not the choice itself; the freedom and responsibility of choice still need to be borne by humans themselves.

This is precisely the significance and purpose of this "information confrontation" pre-competition market.

Only by establishing such a market can we perhaps avoid, to the greatest extent, the monopoly of information by a certain model or a certain company—not only to protect the interests of ordinary people from being eroded by large companies, but even more to prevent the interests of all humanity from being threatened by artificial intelligence at the moment a certain turning point arrives.

In the face of rapidly advancing artificial intelligence, only by keeping the right and responsibility of choice in human hands will humanity ultimately not silently and willingly walk down a road to servitude.