tech
May 1, 2026
Amid Mythos' hyped cybersecurity prowess, researchers find GPT-5.5 is just as good
New results suggest Mythos’ cyber threat isn’t “a breakthrough specific to one model.”
TL;DR
- OpenAI's GPT-5.5 demonstrated performance similar to Anthropic's Mythos Preview on cybersecurity evaluations conducted by the UK's AI Security Institute (AISI).
- GPT-5.5 slightly outperformed Mythos Preview on 'Expert' Capture the Flag challenges, passing 71.4% on average compared to Mythos' 68.6%.
- Both GPT-5.5 and Mythos Preview showed progress on 'The Last Ones' simulation, a complex data extraction attack, though neither model could handle the 'Cooling Tower' simulation of power plant control software disruption.
- AISI suggests that the cybersecurity threat posed by Mythos is likely a result of general AI improvements in autonomy, reasoning, and coding, rather than a model-specific breakthrough.
- OpenAI CEO Sam Altman criticized 'fear-based marketing' and stated that highly capable models will need to be released, albeit carefully.
- OpenAI is controlling the release of its cyber-focused models, such as GPT-5.4-Cyber and the upcoming GPT-5.5-Cyber, to security researchers and critical cyber defenders through its Trusted Access program.
Continue reading the original article