How powerful is Claude Mythos as a hacker really?
The British AI Security Institute (AISI) tested Claude Mythos Preview for cyber capabilities. The model represents a clear leap forward compared to previous AI systems.
In tests, it was able to independently find and exploit vulnerabilities. It executed complete multi-step cyberattacks without human assistance.
In a 32-step simulation (corporate network), it succeeded completely 3 out of 10 times. On average, it completed 22 of the 32 attack steps.
In expert-level CTF hacking tests, it achieved approximately 73% success. The model can significantly accelerate tasks that take humans days. It can autonomously perform attacks on poorly secured systems. However, tests were conducted in simplified environments without real defenses. Therefore, it remains unclear how well it performs against strong, realistic security.
The results primarily show risks for poorly secured networks with access.
AI lowers the costs and effort required to execute cyberattacks.
Evaluation methods must evolve to properly test such powerful AI. Mythos is therefore not yet a “super hacker,” but it is a serious accelerator of cyber threats.
Read the gov.uk article for more information.