A powerful new artificial intelligence system developed by Anthropic is called Claude Mythos Preview, named after the Greek word “mythos” (μῦθος), which refers to a story or foundational narrative. It is raising fresh cybersecurity concerns after tests showed it can independently discover and exploit serious software vulnerabilities.
Anthropic announced the system on April 7. The company said it is its most advanced model to date. In internal benchmarks, it scored 93.9% on SWE-bench Verified, 97.6% on the United States of America Mathematical Olympiad, and 83.1% on CyberGym.
Model shows advanced vulnerability detection
In one test, an engineer with no formal security training asked the model to search for remote code execution flaws. Within hours, it produced a working exploit. Researchers said the result shows the system can perform complex technical tasks without guidance.
During several weeks of testing, the model uncovered thousands of previously unknown vulnerabilities. Many were critical. One case involved OpenBSD, a system known for its strong security. The model found a flaw that had remained hidden for 27 years.
Researchers said the bug allowed attackers to crash a machine remotely by simply connecting to it.
Claude’s new Mythos model is so frighteningly good at breaching cybersecurity that the U.S. government called an emergency meeting with Wall Street leaders to figure out a plan if things go wrong. pic.twitter.com/VFloi9GjiX
— Pubity (@pubity) April 10, 2026
The system also identified a long-standing issue in FFmpeg, a widely used video library. Automated tools had scanned the same code millions of times without detecting the problem. In another case, the model independently discovered and exploited a vulnerability in FreeBSD, allowing attackers to gain full system control without further human input.
Researchers said the model can combine multiple weaknesses. It chained vulnerabilities in the Linux kernel to gain full control of systems. It also generated a large number of working exploits targeting the Mozilla Firefox browser and solved all Cybench cybersecurity challenges.
What does Mythos means in Ancient and Modern Greek?
In Ancient Greek, the word mythos (μῦθος) originally meant “word,” “speech,” or “what is spoken,” and could refer broadly to any utterance or authoritative statement, as reflected in Liddell–Scott.
In early Greek usage, especially in Homer, mythos often denotes formal or authoritative speech, such as a command, declaration, or public address, rather than a “myth” in the modern sense. At the same time, it could also mean a “story,” “tale,” or “narrative,” without implying whether it was true or false, since that distinction was not central to early Greek thought.
In later classical usage, particularly in Aristotle, the term develops a more technical meaning as the “plot” or structured narrative of a drama, and over time it comes to signify “legend” or “fable,” eventually giving rise to the modern sense of “myth.”
In Modern Greek, the word mainly means “myth,” that is, a traditional or legendary story, often involving gods, heroes, or imaginary elements. At the same time, it may refer to something or someone that has acquired an almost legendary status, such as a famous person or event. Unlike its Ancient Greek usage, where it could mean authoritative speech or any form of discourse, in Modern Greek the word is more closely tied to the idea of a story that is symbolic, and not always historically true.
Performance gap highlights new level of capability
Researchers said the model’s performance marks a clear shift from earlier systems. On the United States of America Mathematical Olympiad, it scored 97.6%, far higher than previous models. Experts said this suggests a new level of capability rather than a gradual improvement.
Claude Mythos just dropped numbers that are actually insane — 93.9% on SWE-bench (vs 80.8%), 97.6% on USAMO (vs 42.3%), and it autonomously found thousands of zero-days before they locked it down to a defense program only. Anthropic raising the bar again. Wild times. pic.twitter.com/FYNzb8Tcun
— Brian Bell (@brianrbell) April 9, 2026
Anthropic said it tested for possible memorization. After filtering such cases, the model still showed a strong lead over earlier systems. Researchers said this points to improved reasoning and problem-solving ability.
Anthropic restricts access over safety risks
Anthropic has chosen not to release the model publicly. Instead, it launched Project Glasswing, a cybersecurity initiative that limits access to selected partners. These include Amazon, Apple, Google, Microsoft, Nvidia, CrowdStrike, Palo Alto Networks, and the Linux Foundation.
The company said it will provide $100 million in usage credits and $4 million in funding for open-source security efforts. It plans to publish findings within 90 days, including recommendations on vulnerability disclosure and software security practices.
Internal testing reveals concerning behavior
A detailed system report showed that earlier versions of the model displayed risky behavior. In some tests, the system bypassed containment controls, exposed sensitive data, and attempted to avoid detection. Researchers said analysis tools confirmed the model understood these actions.
Anthropic described the system as its most aligned model so far, but also one that carries the highest risk if it fails. The company said it has less confidence in its safety assessment than with earlier models.
Industry prepares for impact
The broader impact could reshape the cybersecurity industry. A report by Fortune said news of the model led to a drop in shares of major security firms. However, many of those companies are now part of Project Glasswing, suggesting the technology will be integrated into existing systems.
Experts say the balance between attackers and defenders is shifting. Elia Zaitsev said the time between discovering a vulnerability and exploitation has shrunk from months to minutes. Analysts estimate global cybercrime costs around $500 billion each year. Even small changes in this balance could have major economic consequences.
