Anthropic's Claude Mythos outperforms GPT-5.5 at AISI, solves 'Cooling tower'

Technology Oct 02, 2026

Anthropic's Claude Mythos just outperformed OpenAI's GPT-5.5 in cybersecurity tests run by the U.K.'s AI Security Institute (AISI).

Notably, Mythos cracked the tough "Cooling Tower" task that no model had solved before, showing it's getting better at spotting software flaws and handling tricky cyber problems.

AISI cites 2.5 million token limit

AISI says these improvements could really shake up how we defend against cyber threats.

But there was a catch: all models were tested with a 2.5-million-token limit to keep things fair, which might actually hold them back a bit.

As AI keeps advancing fast, the need for bigger, more flexible cyber-testing methods is becoming clearer.