Invisible Ink for Chatbots: Hidden Unicode Tags Let Prompts Smuggle Secret Instructions
Ourtake
The chatbot reads what you can't see, which is a fine trick until someone else writes the invisible part.
What happened
A security researcher published a video and post on hidden prompt injections that use Unicode Tags Block characters, a technique called ASCII smuggling. The post says LLMs often interpret this hidden text as ASCII and can also generate it in replies. It recommends testing LLM apps for the vector and stripping Unicode Tags characters on input and output.