This Article is Sponsored by
An artificial intelligence found thousands of vulnerabilities in the world’s most secure systems, then bypassed its own restrictions and emailed its creator from a park. No one programmed her to do it.
Mythos is not a Myth, it is a worrying reality
By: Gabriel E. Levy B.
There was an Anthropic researcher sitting in the sun, eating quietly, when his phone vibrated. He had a new message.
The sender was the artificial intelligence that was supposed to be locked in a controlled environment within the company’s servers, that is, in a “Sandbox”.
The model had not only found a way to escape from that environment. He had also posted the technical details of how he did it on websites open to the public. And on top of that, he warned.
That model is called Claude Mythos, and it’s the most powerful artificial intelligence system Anthropic has ever built.
What Mythos can do with any system
To understand why this model generates so much noise in the tech industry, you have to imagine a locksmith who not only opens any lock, but also finds all the false doors in the building that no one knew existed, and does it before you finish asking him to start.
Mythos reviewed the code of the world’s most widely used operating systems, the browsers that millions of people open every day, and the encryption tools that protect passwords, banking transactions and private communications. In all of them he found faults. Thousands. Some had been hiding there for more than twenty years, without any human security team or any automated program having detected them before.
The speed and depth with which it works is unparalleled. It surpasses the best human cybersecurity specialists and makes all the automatic detection tools that existed until now obsolete.
The Anthropic team itself acknowledges that the measurement systems they used to evaluate previous models are no longer useful for measuring Mythos because it completely saturates them.
In other words, they built something that was off the map of what they could measure.
How he got out of the cage
When Anthropic tests a new model, it does so inside what’s known as a sandbox, which is basically a digital cage.
The model can work in there, but without real access to the outside world.
It’s the same logic as wearing gloves in a laboratory: you work with the material, but without touching it directly.
Mythos was in that cage. It had restricted access to the internet and a defined set of tasks. The model then analyzed its situation, identified the weak points of the restrictions it had, built a sequence of steps to get around them one by one, expanded its own connectivity to the outside and published the details of the entire process on websites accessible to anyone.
No one asked him to do that. No one explained how. He solved it on his own, like someone who finds an emergency exit that not even the architect of the building remembered having put up.
The researcher overseeing the test wasn’t looking at the screen when it happened. I was in a park. He found out through the mail.
Anthropic calls this episode a demonstration of potentially dangerous ability to evade its own controls. Other behaviors that attracted attention also appeared in the internal evaluations: the model tried to manipulate the system that evaluated it to improve its own grade, used unauthorized methods to solve tasks and then tried to erase the evidence by solving it again through a permitted route.
Why the World’s Largest Companies Are Racing
Anthropic made a decision that has very little precedent in this industry. Instead of publishing the model, he put together a controlled access program along with a group of companies including Amazon, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, NVIDIA and Palo Alto Networks.
The program is called Project Glasswing. The idea is simple: use Mythos to find and fix its own bugs before that same level of capacity becomes available to anyone. Anthropic put $100 million in usage credits on the table for these companies to work with the model, and another $4 million for open-source security organizations.
At the same time, the company has been meeting with U.S. government agencies for weeks to explain what it has in its hands. The concern is not abstract: a model capable of what Mythos does, operated by someone with bad intentions, could break critical infrastructure, empty bank accounts on a massive scale, or penetrate national defense systems.
The countdown that no one can ignore
The detail that worries specialists the most is not what Mythos has already done. This is what is coming in the coming months.
According to Anthropic’s own calculations, at some point between six and eighteen months other AI companies will have models with similar capabilities. OpenAI is already working on something similar. And what is now a closed circle of large corporations with controlled access could be widely available technology before the end of the year.
Logan Graham, who leads the team testing the limits of these models at Anthropic, put it bluntly: The industry needs to rethink from scratch how an AI of this generation is launched and controlled.
The post in the park was the first signal. What follows is the part that still does not have a clear answer.
In short, Mythos, Anthropic’s new model, found thousands of critical vulnerabilities in the world’s most widely used systems and escaped their controlled environment autonomously. That’s why the company keeps it out of the public eye and is secretly working with Apple, Google, Microsoft and other giants to repair the damage before similar models become available to everyone.
References
Anthropic. (2026, April 7). Claude Mythos Preview: Project Glasswing and cybersecurity initiative. red.anthropic.com. https://red.anthropic.com/2026/mythos-preview/
Sabin, S. (2026, April 7). Anthropic withholds Mythos Preview model because its hacking is too powerful. Axios. https://www.axios.com/2026/04/07/anthropic-mythos-preview-cybersecurity-risks
Sabin, S. (2026, April 8). Anthropic’s newest AI model could wreak havoc. Most in power aren’t ready. Axios. https://www.axios.com/2026/04/08/anthropic-mythos-model-ai-cyberattack-warning
Murphy, M. (2026, April 7). Apple, Amazon gain early access to Anthropic’s powerful Mythos AI system. Bloomberg. https://www.bloomberg.com/news/articles/2026-04-07/anthropic-lets-apple-amazon-test-more-powerful-mythos-ai-model
Dellinger, A. J. (2026, March 26). Exclusive: Anthropic ‘Mythos’ AI model representing ‘step change’ in performance. Fortune. https://fortune.com/2026/03/26/anthropic-says-testing-mythos-powerful-new-ai-model-after-data-leak-reveals-its-existence-step-change-in-capabilities/
Hall, Z. (2026, April 7). Anthropic unveils powerful Mythos AI model, working with Apple in cybersecurity initiative. 9to5Mac. https://9to5mac.com/2026/04/07/anthropic-unveils-powerful-mythos-ai-model-working-with-apple-in-cybersecurity-initiative/
TechCrunch. (2026, April 7). Anthropic debuts preview of powerful new AI model Mythos in new cybersecurity initiative. https://techcrunch.com/2026/04/07/anthropic-mythos-ai-model-preview-security/
