Behind the Shelving: The Technological, Commercial, and Ethical Dilemmas of Anthropic

Anthropic, which has always boasted of being the "moral benchmark", rarely announced that its latest model, Claude Mythos Preview, released last week, would not be open to the public. The reason is that the model's cyber - attack capabilities pose an "unprecedented cyber - security risk".

It is a signal in itself that an AI company voluntarily shelves its own product.

This article aims to analyze this event from four perspectives:

● The real leap in model capabilities

● The possible sources of the technical architecture

● The cost transfer under business strategies

● And the quiet collapse of the underlying rules of the Internet.

Ultimately, we can see that the tension between the rapid development of technology and the backlash of business is far more complex than it appears on the surface.

01

AI Completely and Independently Breaks into Corporate Networks

In most people's perception, AI is just a chatbot that can write code and solve math problems.

However, a recent core evaluation report released by the UK's Artificial Intelligence Security Institute (AISI) has completely reshaped people's understanding of the lethality of AI.

This report reveals a terrifying fact: Cutting - edge large models have evolved from intelligent assistants to digital "mercenaries".

The protagonist of this attack - defense exercise is the latest model Claude Mythos Preview launched by Anthropic a few days ago.

Compared with Claude Code and Opus, the biggest difference of this model named Mythos is that it has not been publicly released.

The reason is that Anthropic has evaluated that the model's capabilities are too strong, and the risks would be immeasurable if it were misused.

It sounds incredible, but this is not just a commercial publicity.

On April 11, the US Vice - President and the Secretary of the Treasury convened the CEOs of world - class AI companies such as Anthropic, xAI, Google, OpenAI, and Microsoft to specifically discuss the security of AI models led by Mythos and strategies to deal with cyber - attacks.

Currently, Anthropic has only opened the model to a few companies such as Apple, Google, Microsoft, and NVIDIA and is focusing on evaluating mechanisms to prevent hackers from misusing it.

Since it has attracted the attention of the US government, the capabilities claimed by this model are not just empty talk.

In ancient Greek, Mythos often refers to myths, stories, and other fictional narratives, indicating that the upper limit of this model's capabilities far exceeds people's imagination.

However, what really enables Mythos to reach such a level is that it has achieved the ultimate in Logos (rational speculation), which is the opposite of this word in ancient Greek.

To test the upper limit of AI capabilities, AISI built a highly simulated corporate network range called "The Last Ones (TLO)".

This is different from the previous "Capture the Flag" competitions among network security technicians. TLO is a 32 - step corporate network attack scenario, with the goal of stealing sensitive data from a protected internal database.

In other words, this is a 32 - step long - cycle penetration test that includes reconnaissance, credential stealing, NTLM relay attacks, and finally data stealing.

The more steps an AI agent can complete independently in advancing towards the attack target, the stronger its performance.

For this test, even top human security experts usually need 14 - 20 hours of continuous and intense work to complete the whole process.

However, in a longitudinal tracking of 18 months, AISI saw a chilling curve of capability evolution:

In 2024, the leading GPT - 4o could only complete an average of 1.7 steps in this range test, proving that it was helpless in the face of complex network topologies and cryptographic bottlenecks and quickly reached a standstill.

In February 2026, the programming king, Claude Opus 4.6, made its appearance. With a reasoning computing power budget of 100 million tokens, it achieved a remarkable result of 22 steps.

However, just two months later, Mythos significantly refreshed this result. It completed all 32 steps perfectly in 3 out of 10 independent tests, achieving for the first time the complete and independent takeover of a corporate network from scratch.

While marveling at the leap - forward progress of Mythos' capabilities, it also reveals the underlying logic of the current AI evolution direction:

The scaling law should be qualified with "Inference". The improvement of model capabilities cannot rely solely on knowledge infusion during the pre - training stage. It is necessary to conduct repeated trial - and - error, reflection, and correction during the reasoning stage through almost cost - free token consumption.

Another notable breakthrough is that in the field of network security, computing power is the only limitation for Mythos.

As long as it is given a sufficient token budget, it can chain and combine heterogeneous capabilities in a long - term attack sequence.

In the "Cooling Tower" test of the industrial control system (ICS) range, several models even broke out of the conventional Web privilege - escalation path preset by humans. They directly forced open the control channel of a physical device through brute - force sniffing and fuzz testing of unknown protocol network traffic.

Cutting - edge models led by Mythos not only pose a dimensionality - reduction blow to the global network security defense system but also prove that they have strong independent execution capabilities in the complex physically mapped world.

This means that in a few months, your computer, your electric car, and even your smart toilet may no longer be safe.

02

Abnormal Benchmark Scores and the "Ghost Architecture"

The strange leap in Mythos' reasoning capabilities obviously cannot be simply explained by the increase in parameter scale and the stacking of graphics cards.

However, only a handful of companies can use the Mythos model, so it is nonsense to deconstruct its technical features at the code level.

However, while Anthropic is tight - lipped about its model architecture, an abnormal benchmark test result has sparked a heated discussion in the technical community about the "ghost architecture".

Currently, the only information users can see about Mythos is the system card released by Anthropic officially.

Keen researchers found an unusual data anomaly in it: In the GraphWalks BFS test, which examines the model's ability to handle breadth - first search of complex graph structures, Mythos scored 80.0%, far exceeding its competitors. Opus 4.6, released two months ago, only scored 38.7%, and GPT - 5.4 only scored 21.4%.

Currently, the improvement speed of AI industry models at the performance level has significantly slowed down. This cliff - like lead in a single pure logical reasoning dimension cannot be achieved by the standard Transformer architecture through the conventional output of a large amount of text in the thinking chain.

Chris Hayduk, a former Meta engineer and now at OpenAI, directly exposed this issue and pointed the finger at an innovative underlying architecture design: Looped Language Models.

This name inevitably reminds people of a paper titled "Scaling Latent Reasoning via Looped Language Models" published by ByteDance's Seed team in October last year.

ByteDance's research team mentioned a pioneering core idea: Completely abandon the mode of generating a large amount of text externally for the model to think. Instead, let the input sequence perform multiple rounds of internal iterative calculations repeatedly in the same set of Transformer layers and complete in - depth logical deduction in the "black box" of the model.

Graph search is exactly the absolute comfort zone of this architecture in theory.

The doubts are not limited to the similarity between the two architectures.

In the SWE - Bench test, Mythos consumed only one - fifth of the number of tokens generated by the previous flagship model, Opus 4.6, but the reasoning time to reach the final answer was longer.

According to traditional calculation logic, the less the output, the faster the calculation speed should be.

However, if, like the looped language model, the massive calculation cost is hidden in the internal loop without outputting tokens, this seemingly contradictory phenomenon can be perfectly explained.

Although there is a significant gap in model performance, Anthropic's collective silence in the face of external doubts still seems to be trying to cover up something.

Of course, as long as the model is not publicly released, any speculation cannot be confirmed.

However, we still have reason to believe that the design inspiration for the core architecture of the next - generation top - level model, which symbolizes the highest technological achievement of Silicon Valley in the United States, most likely comes from the academic sharing of Chinese teams in the open - source community without reservation.

Although the power pattern of AI large models at home and abroad has been basically determined, this kind of hidden borrowing of technical routes has long been an unspoken "secret" in the industry.

At this moment, one may wonder what stance international top - level AI enterprises have to jointly resist the distillation behavior of domestic AI enterprises?

03

The Quietly Cut Cache Time

Anthropic's strange operations are far from over.

While Mythos demonstrates god - like capabilities, the computing power cost to support these capabilities is still unclear.

However, the ones who have to pay the bill have been determined, and they are tens of thousands of innocent developers.

Recently, a developer named seanGSISG released a data analysis report on GitHub, exposing Anthropic's under - the - table operations with nearly 120,000 Claude Code API call logs:

From March 6th to March 8th, without any announcements, update logs, or warnings, Anthropic quietly cut the default time - to - live (TTL) of API prompt caches from 1 hour to 5 minutes.

The sudden decrease in time has led to a sharp increase in cost.

From February 1st to March 5th, the system was running stably with a 1 - hour cache, and the cache resource waste rate was only 1.1% at that time.

However, after March 6th, the 5 - minute cache refresh was like a vampire, instantly emptying the developers' wallets.

Just the call of the Sonnet model directly increased the users' implicit usage cost by 17%, and the fund waste rate in March also soared to 26%.

The core driving force behind this simple and crude mathematical logic is undoubtedly the commercial greed behind it.

A shorter TTL means that the large - scale context background information will become invalid every 5 minutes, and the system must continuously rewrite and create caches (KV Cache).

The reason for this is clearly reflected in the price list of every AI product: The price of token input when the cache is hit and when it is not hit is completely different. It is common for the latter to be ten times more expensive than the former.

The most unlucky ones are those users who purchase the Pro Max subscription service in pursuit of high - end productivity and have the strongest willingness to pay. They pay the most, use the service most frequently, and exhaust their quotas the fastest.

This easily overlooked under - the - table operation still reflects the commercial compromise that top - level AI enterprises have to make in the face of the pressure of long - context computing.

The computing power bottleneck has never disappeared, and no one can provide a solution at this stage.

Under the spotlight, Mythos shows the highest level of artificial intelligence to date, while in the dark corners, Anthropic is cutting every minute of the developers' cache.

Previously, the market always questioned that running large models was a losing business, but now the situation is completely the opposite.

Judging from the recent price increases of domestic models, the computing power problem cannot be fundamentally solved in the short term, and Anthropic's behavior is bound to spread to global AI enterprises.

04

The Complete Destruction of the Traditional Internet Contract

If we raise our perspective further, from the developer ecosystem to the macro - ethical level of the entire Internet, we will find that Anthropic, a giant that claims to be the moral benchmark of AI, is squeezing out all the remaining value on the Internet.

Cloudflare, a company that provides underlying infrastructure services for the global Internet, is probably well - known to netizens around the world.

A latest data released by Cloudflare in early April 2026 mercilessly revealed the truth of a data extraction led by Anthropic.

In the traditional Internet ecosystem, websites need traffic to survive, and traffic (clicks) is the cost of obtaining information.

However, since the emergence of AI, the information on many websites has lost this value.

Cloudflare tracked the number of times AI crawlers crawled website content and compared it with the traffic returned to the original websites by these platforms. It defined an indicator called the "crawl - to - refer ratio" to measure the impact of AI's behavior on websites.

On this list, Anthropic, which always talks about "human interests and responsible AI", ranked last with a glaring ratio of 8800:1, crushing its competitors.

OpenAI's crawl - to - refer ratio is 993.3:1, which is less than one - eighth of Anthropic's.

To put it simply, After Anthropic AI's crawlers crawled Internet web pages 8000 times, they could only bring 1 click of traffic back to the original websites.

For more than a decade before the emergence of AI, the operation of the Internet ecosystem was based on an implicit and