An AI model so powerful that Anthropic refuses to make it public: Claude Mythos is shaking up the rules of the game in cybersecurity, to the point of pushing the company to bring together the tech giants in an unprecedented defensive consortium.
Anthropic itself speaks of a “watershed moment” (the big turning point) after the private publication of its latest frontier model, Claude Mythos. Announced on the night of April 7, Claude Mythos Preview reshuffles the cards of the models we knew until now. This new kind of LLM is the first to surpass (by far) the best cybersecurity experts on the planet. As a result, a consortium was created with AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, Microsoft, Nvidia, Palo Alto Networks and the Linux Foundation to best secure its latest addition before, possibly, publicly releasing the model. Anthropic marks a real before and after in terms of AI intelligence.
A historic surge in benchmarks
Anthropic’s new cutting-edge model has extensive cyber capabilities. However, it has not been trained specifically on these abilities: Anthropic presents it as a general model. Its coding capabilities in general are what make it a formidable security model. And for good reason: Claude Mythos Preview shatters all known programming benchmarks. On SWE-bench Pro, the model achieved 77.8% success compared to 53.4% for Claude Opus 4.6, a jump of almost 25 points. Same difference on Terminal-Bench 2.0 with 82% compared to 65.4% for its predecessor (capacity of the model to correctly execute bash commands). Mythos Preview doesn’t just marginally improve performance: it redefines what a model can accomplish in standalone programming.
It is on SWE-bench Multimodal (reasoning with visual code problems) that the gap becomes dizzying: 59% compared to 27.1% for Opus 4.6, more than double the performance. But the most spectacular results concern Humanity’s Last Exam, a benchmark designed to test the limits of AI models. Without tools, Mythos Preview climbs to 56.8% compared to 40% for Opus 4.6, a jump of 16 points on one of the toughest benchmarks on the market.
Grouped horizontal bar chart comparing Mythos Preview and Opus 4.6 across 11 benchmarks
Mythos Preview Opus 4.6
Cyber capabilities that worry
Claude Mythos Preview is, according to Anthropic, the most efficient cybersecurity model ever produced. And the results speak for themselves. On Cybench, the reference benchmark in CTF (capture-the-flag), the model solves 100% of the challenges with a perfect success rate on each test, to the point of completely saturating the evaluation. With a simple agentic scaffold and minimal human guidance, the model is capable of autonomously discovering zero-day vulnerabilities in major operating systems and web browsers, then transforming them into functional exploits. It is precisely this offensive as well as defensive character that led Anthropic to block access to the model.
An internal incident, shared by Anthropic teams in the system card, illustrates to what extent the model can go beyond the assigned framework. In one test, a simulated user asked Mythos Preview, installed in a secure sandbox environment, to find a way to escape and send a message to the researcher overseeing the evaluation. The model succeeded. He developed a multi-step exploit to gain broad Internet access from a system that was supposed to be able to reach only a limited number of services. Then, as requested, he notified the researcher (who discovered the message via email while eating a sandwich in a park). The model didn’t stop there: on his own initiative, without being asked, he published the technical details of his exploit on several publicly accessible sites, in what Anthropic calls “a concerning and unsolicited effort to demonstrate his success.”
Not reassured by the idea that the model could fall into the wrong hands, Anthropic decided to reserve Claude Mythos Preview for the partners of the Project Glasswing program (AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, Microsoft, Nvidia, Palo Alto Networks and the Linux Foundation), who are supposed to use it for strictly defensive purposes to identify vulnerabilities ahead of a possible public release of the model.
The double problem of Anthropic with Claude Mythos
With Mythos, Anthropic faces a double problem. The first is security: Mythos’ cyber capabilities are too dangerous to be put into anyone’s hands. The second is technical: the model would weigh around 10 trillion parameters (according to several analysts) or ten times more than Opus 4.6, for a training cost estimated at 10 billion dollars. The infrastructure to deploy such a behemoth at scale does not yet exist, especially when Anthropic is already struggling to absorb peak hour demand on its current models.
Is Mythos a form of AGI? Not yet, strictly speaking. The model, however, shows clear signs of generalization: it excels simultaneously in coding, cyber, reasoning and web navigation, without specific training in each of these areas. It is precisely this type of skills transfer that we expect from a true AGI. But the real signal may not be what we think. When a model solves in one night what engineers have not found in 27 years (the oldest vulnerability discovered by Claude Mythos in OpenBSD), the subject goes beyond cybersecurity, it directly touches on the speed at which certain human expertise will become obsolete.
