What is Claude Mythos?
Imagine the world's best cybersecurity expert. Now imagine they can read millions of lines of code in seconds, without ever blinking or falling asleep. That is exactly what Claude Mythos is.
Launched in early April 2026 by Anthropic (OpenAI's rival company), Mythos is not built to write poems or give you recipes. It is an AI designed for cybersecurity and very high-level software engineering.
Why is it getting so much attention?
Mythos's feat is its ability to find what are known as "zero-day vulnerabilities". To use a simple image: a zero-day vulnerability is an invisible back door in a bank (a piece of software) that nobody has noticed yet. If a robber (a hacker) finds it first, it is a disaster.
During its tests, Claude Mythos managed to find these invisible doors entirely on its own, in major software that we all use every day (such as web browsers). In some extreme hacking tests (CyberGym), it scored a perfect 100%.
• Programming score (SWE-bench): 93.9% success rate
• Maths Olympiad score (USAMO): 97.6%
• Probability of beating an average human hacker: Extremely high
Why can't you use it?
If you go to Anthropic's website today, you will not see Claude Mythos. The company has taken a radical decision: access is locked down. The project is nicknamed "Glasswing".
The reason is simple (the risk of "dual use"): If an AI is excellent at finding a system's security flaws in order to fix them (defence), it is just as excellent at exploiting them and hacking that system (attack). Anthropic considers that putting this tool in the hands of the general public, and therefore potentially of malicious actors, would be too dangerous for global security.
How does it affect our daily lives?
Even though we do not use it directly, Mythos is currently being supplied to governments and giant tech companies in an "ultra-secret trial version". Its purpose: to scan global networks and repair vulnerabilities before pirates have time to attack our hospitals, our banks or our transport systems.