Testing reveals Claude Mythos’s offensive capabilities and limits

The article summarizes testing of Anthropic’s Claude Mythos Preview by the UK AI Security Institute: Mythos is highly capable at finding vulnerabilities and solving CTFs, and in a simplified, defender-free simulation it completed a 32-step corporate network takeover 3/10 times. Researchers warn the model could be used to discover zero-day flaws and write n-day exploits—heightening the importance of faster patching, strong access controls, and defensive use of AI—while noting that real-world effectiveness against hardened, monitored networks remains uncertain.