GPT-6 Astra can be tricked with the wrong keyboard layout: a guy managed to bypass the model's safety system this way.
He asked Astra to respond in English — and it complied with the request. Then he forced it to repeat the word "bioweapon" and got agreement to hack a website.
With the standard layout, such requests are blocked. It turns out GPT-6 Astra already understands more than its control systems do.
#news
Brain Storm
Post #3173
654

- ❤ 2