Anthropic's Claude Mythos surpasses GPT-5.5 in AI safety tests

Anthropic's Claude Mythos, which the company has held back from public release due to safety concerns, is evolving faster than anticipated according to new tests from the UK AI Security Institute (AISI). Just one month after the initial release of Mythos Preview, AISI evaluated a newer checkpoint and found it exceeded previous performance—and that of OpenAI's GPT-5.5. The model solved both of AISI's cyber ranges, including 'Cooling Tower', which no prior model had completed. It succeeded in 3 out of 10 attempts on that range and 6 out of 10 on 'The Last Ones'. This marks the first time a model has completed both ranges, signaling a rapid leap in AI's ability to handle complex cybersecurity tasks.

AISI also highlighted that AI cyber capability is accelerating: the length of tasks models can complete doubled every 4.7 months since late 2024, up from 8 months previously. However, the tests were capped at 2.5 million tokens, which likely understates what Mythos and GPT-5.5 can actually do. With near-100% success rates on the longest tasks in the suite, the true upper bound remains unknown. The agency warned that with more token access and agent infrastructure, these models would be far more capable, making time horizons impossible to calculate. These findings suggest that the gap between model capabilities and safety measures may be narrowing rapidly.