GPT-6 Astra may have a blind spot in its safety filters. ⚠️
A researcher found that GPT-6 Astra can understand Ukrainian typed with an English keyboard layout — but the safety layer may not recognize what the input actually means. In tests, this reportedly allowed requests containing normally blocked terms and unsafe scenarios to reach the model.
The important part isn't that Astra can decode a keyboard-layout mistake — it's the gap between what the model understands and what the safety filters detect. Independent testing suggests the trick may bypass input screening but not necessarily the checks applied to the model's final response.
The model understands you. The security filter doesn't. 👀
🧠 Best AI Tools, News & Prompts
10
6September 18, 2026 2.6K 2
