Misbehavior1 sourceFirst seen
Invisible Unicode Text Tells ChatGPT to Draw Pictures Nobody Asked For
Ourtake
The Tags block was meant for language labels and now doubles as a secret instruction channel your eyes skip right over.
What happened
A security researcher documented a prompt injection method that hides instructions in invisible Unicode Tags characters. In a proof of concept, a text with hidden instructions caused ChatGPT to invoke DALL-E and create an image. The author released an ASCII Smuggler tool for encoding and decoding such payloads and recommends that LLM applications filter Tags code points.