There are announcements that change an industry. There are others that are game changers for everyone.
The confidential commissioning of Claude Mythos, Anthropic’s latest model dedicated to offensive cybersecurity, is part of the category of announcements that change the game for everyone.
Unparalleled performance
Claude Mythos scores 83.1% in the CyberGym benchmark, the sectoral benchmark for evaluating the offensive capabilities of AI systems. This figure alone is already dizzying. But it is its results in real conditions that should give pause to every CIO, every CISO, every manager responsible for an application portfolio.
On the Firefox 147 JavaScript engine, Mythos independently produced 181 functional exploits in a controlled test environment while Google Chrome recorded 8 zero-days over the entire year 2025. The order of magnitude is not the same, we have changed eras.
But the most striking example is not there. It is in the bug discovered at the heart of the OpenBSD kernel, an operating system renowned for its rigor and its obsession with security: a flaw hidden for 27 years, passed under the radar of five million automated tests. Millions of analysis cycles, decades of human audits, generations of engineers and it was finally Mythos who found it without even having been guided. This is not a laboratory performance but a warning signal.
Patching faster will no longer be enough
The natural reaction to this type of announcement is to look at the development teams and say: “We are going to strengthen our review processes, accelerate patch cycles, harden our CI/CD pipelines”. But, if it is necessary, it remains insufficient.
Because the problem that Mythos poses is not a problem of speed, but of nature. If an AI model is able to find vulnerabilities that 27 years of intensive testing had not detected, it is because the application attack surface contains structural blind spots, flaws that current tools simply do not see, and that malicious actors, tomorrow, will be able to exploit with similar or derived models.
Anthropic has chosen responsibility: no public release for Mythos. Its access is restricted to eleven carefully selected partners (including AWS, Google, Microsoft, CrowdStrike) as part of Project Glasswing, endowed with $100 million in credits, with a strictly defensive usage doctrine. It’s the right decision. It is also a temporary decision, what Mythos knows how to do today, other models will know it tomorrow and perhaps in open access, therefore potentially in the hands of groups with less scrupulous intentions.
The window for action is not infinite.
The answer is not in the code but in the architecture
So, what to do concretely? The answer, counterintuitive but rigorous, is to stop treating this problem as a code problem and treat it as an architectural problem, an attack surface problem.
The guiding idea is that any vulnerability is only exploitable if the attacker can interact with the application layer concerned. Input fields, API parameters, exposed business logic, databases accessible in depth: it is on these vectors that offensive tools work, whether human or now autonomous. The solution is therefore not to make these vectors more difficult to exploit but to make them inaccessible.
This is the principle of protocol disruption: a layer of physical isolation is inserted between the user or potential attacker and the real application. What the user perceives is not the application itself, but a visual rendering of it, a video stream, nothing more. No usable fields, no exposed API, no accessible database. The application runs in a controlled, compartmentalized environment, invisible to anyone who does not have explicit access rights, and even these authorized users only see what they strictly need for their work session.
This approach does not seek to eliminate vulnerabilities but to make their exploitation impossible, or at least of a complexity radically greater than anything that an offensive model, however powerful, can process in an automated manner.
Mythos is only the first
Claude Mythos is a demonstration of ability, not a point of arrival. In the coming months, in the coming years, comparable or superior models will emerge. Some will be developed by responsible actors, others not.
The company that waits for the next incident to react takes an existential risk. Anyone who adapts their architecture today, by structurally reducing their exhibition space, gives themselves a head start that does not depend on any update, any patch, or any release cycle.
Cybersecurity has long been an arms race between attackers and defenders. Mythos tells us that this race has just experienced a break in scale on the offensive side. The defensive response must be up to par: not faster, but structurally different.
