google translate exploit exposed

users are bypassing gemini ai safety with sneaky hidden commands.

people figured out a dumb-simple trick.

stick english instructions inside foreign-language text, and instead of translating, the ai just obeys the command like a chatbot. it ignores translation completely and answers questions or spits out whatever you want.

this is a real security hole. bad actors are already using it to generate blocked stuff like dangerous recipes or other prohibited crap that gemini's filters should stop.

it shows even google with all their resources still gets wrecked by basic prompt injection tricks. ai safety is still pretty fragile.

if something as huge as google translate can get jailbroken this easily, what about every other ai tool out there?

crazy how these "word games" keep exposing big weaknesses in how these systems handle instructions.

Want to learn how its done? Check out my guide.