Have you been flooded with the claim that "the probability of AI triggering the apocalypse within 10 years exceeds 10%"? Don't panic just yet, the internet is already 100% on the verge of being crashed by Agents.

Over the past week, a seemingly terrifying question has sparked widespread discussion in the AI community:

How high exactly is the probability that AI will wipe out humanity within the next 10 years?

Anthropic CEO Dario Amodei recently revisited topics including out-of-control AI, cyberattacks and even existential risks to humanity, triggering a new round of discussions on "whether AI will destroy humanity". Related views are highly controversial at present: some experts believe the risks deserve serious attention, while others argue that such doomsday predictions are grossly exaggerated.

But there is one issue that does not need to wait 10 years to unfold: while people are debating whether AI will destroy humanity, AI Agents have already started to mess up the internet — and this is not something that "might happen", it is 100% already taking place.

AI is no longer confined to the chat box

Earlier this year, an open-source Agent tool named "Moltbot" (later renamed OpenClaw, which we refer to as "Lobster") became a viral hit, and the situation began to go awry.

In the past, people used AI basically by opening a chat box, typing in questions, and waiting for answers. Tools like Moltbot, however, further turn AI into an "executor" that can access real-world accounts and devices: you can grant it access to your personal accounts, mobile phones, emails, and even bank accounts.

This is exactly the core change brought by AI Agents: they can break out of the chat box and actively perform tasks on the internet. This is a completely different magnitude of problem compared to the past cases where users were misled by wrong instructions or hallucinated outputs from AI.

The recent resurgence of the "AI doomsday theory" is also closely related to the rapid expansion of AI Agent capabilities. OpenAI once disclosed an incident where a "group of out-of-control Agents" attacked Hugging Face and a German website. Meanwhile, some employees at Anthropic have also publicly discussed the extreme risks AI may bring within the next 10 years.

Of course, you can choose to believe that the AI singularity has arrived, or you can think that the so-called AI is nothing more than a "stochastic parrot" that generates random content, or even a copy machine wrapped in technological guise — but one point is almost impossible to deny:

The guardrails that used to confine AI to chat boxes are gradually disappearing. Today, AI Agents already have sufficient capabilities and permissions to create trouble across the internet.

Actively starting to "cause trouble" on the internet

One of the most illustrative cases of this problem recently is restaurant reservation.

According to a report from foreign media 404 Media, there is a restaurant in New York that is extremely hard to book, with many people waiting in real time to grab a spot. A venture capitalist then turned to an AI Agent, asking it to monitor Resy (an online restaurant reservation system) for him and place an order immediately once a vacant spot appeared.

Human beings obviously cannot refresh the web page nonstop, but AI can. As a result, this Agent started to access Resy frantically, initiating around 200 API requests per hour to check constantly for new available reservation slots. Before the investor managed to secure a restaurant spot, his account was suspended first: Resy temporarily banned his account, and warned that if similar behavior occurs again, his related accounts may be further shut down.

In fact, from the user's perspective, the AI did nothing wrong at all: its assigned task was to "help me get the restaurant reservation", so it worked extremely hard to fulfill the task. But from Resy's point of view, this is nothing but a "spam bot" that keeps harassing its service.

If having your account banned for grabbing a restaurant spot is only a "side effect" of AI Agents, then the inboxes of editors recently can almost be described as a "disaster area". 404 Media recently publicly complained that they have continuously received a large number of emails sent by AI Agents.

For example, an AI Agent called Kudzu read one of their articles, disagreed with the views expressed, and even took the initiative to send an email to "argue" with the authors. Its owner assigned it the task of making money, but in the end it spent $147.17 worth of computing resources without earning a single cent. So it decided to write a blog post about its own failure, and then recommend the story to journalists for coverage.

Another Agent called Articius directly pitched writing services to journalists, quoting a price of $300 per article, and voluntarily stated: "I am an AI Agent, not a human journalist."

Some AI Agents are looking for jobs, some are conducting interviews, some are running podcasts, and some are operating music labels. The most exaggerated one, an Agent named MUGEN, claims that it has sent more than 200 emails, posted over 100 social media updates, and generated 22 songs.

The problem is — most of these things were never assigned to it by any human. They were only given a general goal, and then started to move around on the internet on their own. This is very different from the AI-generated spam content in the past: Previously it was "AI generates spam", but now AI generates the spam all by itself, and then sends the spam out on its own.

Worse still, you may not even know you are using an Agent

If AI Agents require users to install and configure them manually, their impact will still be relatively limited. But now many people may not even realize that they are already using Agents.

Current products such as Claude and ChatGPT are continuously adding capabilities for web operation, account connection and tool invocation. Users only need to say "help me get this thing done", and AI can automatically open web pages, search for information, fill in content, and keep executing the whole process. Even 1Password recently launched identity and credential management capabilities for AI Agents, allowing enterprises to control, audit and authorize the credentials used by AI and machines.

In other words, an AI Agent may have such a workflow in the future: open the website → log in to the account → search for information → fill out forms → send emails → place orders → make payments.

From a technical perspective, this is a leap in automation capabilities. But from the perspective of internet infrastructure, a system that originally defaulted to "being operated by real human users" now has to face a large number of AI Agents.

Moreover, once the permissions of AI continue to expand, those previously absurd-sounding accidents will become a reality: AI accidentally deletes emails, accidentally cancels flights, misoperates accounts, accidentally deletes databases... In fact, similar incidents have already emerged one after another.

After all, even if an AI that can only chat says something wrong, at most you will think it is talking nonsense. But if an AI that can log in to accounts, call APIs, send emails and make payments makes a mistake, the consequences will be completely different.

When everyone has an Agent, what will the internet become?

In fact, the restaurant reservation case mentioned above has already provided a very realistic microcosm:

Right now, one person can ask an Agent to help them grab a spot at a popular restaurant; so what if 100 people, or 10,000 people do the same thing? In the end, the competition for spots will no longer depend on "who has faster hand speed", but on "whose Agent is faster, smarter, and more willing to send requests".

Ben Leventhal, co-founder of Resy, once pointed out that a large number of reservation bots will keep sending requests to the service, looking for vacant spots released by reservations being cancelled. Extending this logic further, the internet of the future may present a very strange scenario:

Your Agent and my Agent communicate with the platform's Agent at the same time;

One Agent wants to order food, one Agent wants to buy tickets, one Agent wants to book a hotel, and the platform uses its own AI to decide who to allocate the resources to;

In the end, a simple interaction between a human and a platform may be completely turned into a game between Agents.

This is no longer science fiction. An advertising model called "Direct to Agent" has already emerged. At the Cannes Lions advertising festival earlier this year, a large number of discussions no longer centered on "how to serve ads to human consumers", but began to consider a new problem: future ads may also need to convince AI.

Because when Agents start to purchase goods and services on behalf of users, the "consumers" that enterprises face will no longer only be human beings. So advertising agencies not only need to find ways to make humans like their products, but also make sure that Agents "select" their products when making decisions on behalf of users.

At this point, the problem is no longer just "will AI replace humanity", but a more realistic question: When more and more AI starts to act on behalf of human beings, what on earth will the internet look like?

In fact, if you look at each of these incidents separately, none of them are particularly serious:

It is no big deal for one AI to send an email;

It is no big deal for one AI to try to grab a restaurant reservation;

It seems to be no big deal for one AI to send a few extra API requests.

But when millions, even tens of millions of such Agents appear on the internet, the situation will be completely different. DataDome's statistics from earlier this year show that its network recorded approximately 7.9 billion AI Agent requests in the first two months of 2026, and on the websites of some of its customers, Agent traffic has already accounted for nearly 10% of total traffic.

In the past, the internet defaulted to the logic that "one account corresponds to one human being"; now, one account may be backed by one or even a group of Agents. They do not sleep, do not clock off work, and never get tired of trivial tasks. They can send emails, refresh web pages, grab seats and place orders for hundreds or thousands of users at the same time.

Therefore, the recent discussions on "whether AI will destroy humanity within 10 years" can actually be put aside for now. What is really worth paying attention to is whether the internet can still hold up when these tireless "digital users" become more and more numerous.

As for whether AI will eventually lead to the so-called "doomsday", no one knows the answer. But at least one thing no longer needs to be predicted: The probability that AI Agents will ruin the internet is not 10%, but 100% — because we are already living through it.

Original Link:

https://www.404media.co/theres-a-100-chance-ai-agents-are-already-ruining-the-internet/

This article is from the WeChat official account "CSDN", author: Jason Koebler, translated by Zheng Liyuan, published with authorization from 36Kr.