Updated
Updated · POLITICO · May 24
UK Institute Finds Mythos Seizes Corporate Networks in 60% of Tests, GPT-5.5 in 30%
Updated
Updated · POLITICO · May 24

UK Institute Finds Mythos Seizes Corporate Networks in 60% of Tests, GPT-5.5 in 30%

1 articles · Updated · POLITICO · May 24

Summary

  • Mythos fully took over a corporate network in 6 of 10 UK AI Security Institute tests, while GPT-5.5 succeeded in 3 of 10.
  • Those results underscore how quickly frontier models' cyber capabilities are improving, with British AI Minister Kanishka Narayan saying the advance is faster than expected.
  • Broadcom, which tested Mythos against its own software code, called the findings "jolting" and said the model was uncovering issues unlikely to have been found by human researchers alone.
  • The same capability could help defenders catch bugs before software is released, but officials and researchers warn it may give attackers and state-backed hackers the bigger near-term advantage.
  • China's push to copy U.S. AI through large-scale distillation attacks is adding to fears that rival states could soon field similarly powerful cyber models.

Insights

With AI automating cyberattacks, can human-led defenses possibly keep pace?
Is China's open-source AI strategy a power play or a move that recklessly arms the world?

AI Models Surpass Human Experts in Cyberattack Simulations: Anthropic Claude Mythos and OpenAI GPT-5.5 Achieve Breakthroughs in 2026

Overview

In April 2026, the UK AI Security Institute reported a major breakthrough in AI's ability to carry out cyberattacks autonomously. Anthropic’s Claude Mythos Preview became the first AI to fully complete a complex corporate network attack simulation, solving tough challenges like 'The Last Ones' and the previously unsolved 'Cooling Tower.' OpenAI’s GPT-5.5 also showed similar high-level performance. These results mark the first time any AI has mastered both of AISI’s most difficult cyber ranges, highlighting a rapid and significant leap in AI-driven offensive cybersecurity capabilities.

...