Anthropic's Mythos is Evolving Faster Than Expected, Reports AI Safety Agency
ZDNET, Thursday, May 14th, 2026
Anthropic's Claude Mythos model has advanced beyond expectations, outperforming earlier versions and GPT-5.5 just one month after release.
Anthropic's Claude Mythos model, which the company considers too powerful for general release, has already demonstrated new capabilities according to testing by the UK AI Security Institute. The newer Mythos Preview checkpoint achieved several firsts in cybersecurity testing, successfully completing previously unsolved cyber range tasks that earlier versions and OpenAI's GPT-5.5 could not.
AISI's findings suggest that AI capabilities in cybersecurity are advancing at an accelerating pace, with the capacity to handle cyber tasks doubling approximately every 4.7 months since late 2024. However, researchers note significant unknowns in their testing methodology, including token limits that may understate the true capabilities of frontier models.