Anthropic’s Mythos Preview was touted as a cybersecurity game-changer, but new tests reveal that OpenAI’s GPT-5.5 is no slouch in the field either. Both models performed similarly on key tasks, from reverse engineering to data extraction attacks.
The UK’s AI Security Institute (AISI) evaluated both AIs using 95 Capture the Flag challenges designed to test their cybersecurity capabilities. Notably, GPT-5.5 solved a particularly tough task in just over ten minutes, whereas Mythos Preview needed a bit more time. Yet, when it came to simulating an attack on critical infrastructure, both faltered.
‘GPT-5.5 passed with flying colours where others have failed,’ said AISI’s lead researcher. ‘But the bar is high in this domain.’ The results suggest that while GPT-5.5 and Mythos Preview are formidable, they still face significant challenges in real-world cybersecurity scenarios.
As A.I. continues to evolve, these tests highlight the ongoing race between innovation and security. For now, it’s a close call, but the stakes couldn’t be higher.
