A Crack in Pandora's Box: What Do We Learn From the Unauthorized Access to Anthropic's Mythos Model?

The model that was meant to protect the network found itself in the wrong hands. Although the damage this time was minimal, this incident is a glaring warning sign for anyone involved in securing artificial intelligence models.

In recent weeks the AI world has been buzzing about the announcement of Claude Mythos Preview, the most advanced model from Anthropic. This is not just another chatbot that will write marketing emails for you; Mythos is a dedicated information security model with exceptional "red" (Offensive) capabilities.

Anthropic, understanding the risk inherent in such a model, chose not to release it to the general public. Instead, it set up Project Glasswing - a limited coalition of strategic partners such as Apple, Amazon, Google, and Microsoft, with the aim of using the model for defensive purposes only: finding zero-day vulnerabilities and strengthening critical infrastructure.

But on the very day of the announcement, it turned out that the wall was not high enough. A group of users managed to gain access to the classified model, and this incident teaches us an important lesson about the gap between security theory and reality on the ground.

Contrary to what it might seem, this was not a complex breach of the company's servers in a "Mission: Impossible" style. Access was obtained in much more prosaic ways:

1. •URL Guessing: The users exploited the fixed format of Anthropic's API addresses to locate the testing environment of the new model.
2. •Failure in an external vendor's environment: The main access was made possible through a work environment of an external contractor (Third-party contractor) of Anthropic. This is a classic failure in the Supply Chain, where the company's own security can be excellent, but the weak link is at one of the partners.

According to reports in Bloomberg and TNW, the group of users used the model mainly to build code for simple websites and to provide live proof of their access, and not to carry out cyberattacks. Even so, the very fact that a model capable of finding thousands of critical security vulnerabilities in operating systems and browsers was accessible without supervision is a serious event.

To understand the scale of the incident, you have to understand what Mythos can do. Unlike previous models, Mythos displays exceptionally high "agentic" capabilities (Agentic capabilities). In tests by the UK AI Security Institute (AISI), the model managed to:

• •Find zero-day vulnerabilities in Linux and OpenBSD systems that survived decades of human review.
• •Chain (Chain) several vulnerabilities together to carry out a full system breach.
• •Write working exploits (Exploits) within just a few hours.

This capability makes it a double-edged sword: it is the best tool for defending the network, but also the most dangerous weapon if it leaks into the wrong hands.

The Anthropic incident highlights three critical points that every organization relying on large language models (LLMs) must take into account:

1. •Securing the model is not only securing the code: Access to the model is the most valuable asset. If you use internal models or models with high permissions, access control must be based on Zero Trust, even inside the organization and certainly toward vendors.
2. •Third-party risks are the new reality: Anthropic learned the hard way that your partners are an integral part of your attack surface. Investing in security Due Diligence for vendors is not a recommendation, but a necessity.
3. •The gap between defense and offense is narrowing: Models like Mythos lower the barrier to entry into the offensive cyber world. What once required a team of expert hackers for months can be done today by a model overnight. The answer to this must be adopting AI-based defense tools at a higher speed.

The Mythos case is a wake-up call. It proves that even the companies most focused on Safety struggle to control the technology when it meets the complexity of the global supply chain. As models become more powerful, the challenge will be less about "what the model can do" and more about "who can operate it".

In a world where artificial intelligence is becoming the main engine of code development, we must adopt new standards of Governance and security, to make sure that the tools meant to protect us are not the ones that bring down the walls.

Relevant links:

• •Anthropic's announcement of Project Glasswing
• •Bloomberg report on the unauthorized access
• •The UK AI Security Institute's report on advanced models