AISI finds Claude Mythos cracks 'Cooling tower' and outperforms GPT-5.5
Technology
Anthropic's new AI, Claude Mythos, just outperformed its earlier results and OpenAI's GPT-5.5 in key cybersecurity challenges, according to the U.K.'s AI Security Institute (AISI).
It nailed "The Last Ones" test six times out of 10 and finally cracked the tricky "Cooling Tower" challenge three times, a first for any model.
AISI: 2.5 million token cap limits performance
AISI pointed out that all models were tested with a 2.5-million-token cap, which might actually hold back top performers like Mythos.
They believe giving these AIs more room to work could make them even better, especially as smarter models are starting to take on bigger roles in cybersecurity.