Analysis
Invisible Prompt Injection — Zero-Width & Unicode Tag Hidden Text in LLMs
Publication date unknown · Discovered stegzero.com
Page published
Publication date unknown · First observed: 4 Nov 2026
Coverage timeline
Single-source analysis — one report is available.
Why it matters
Invisible Unicode payloads can hide instructions in content consumed by LLM applications, making visual review and simple keyword filtering unreliable security boundaries.
StegZero's explainer compares invisible Unicode techniques used to conceal prompt injections or evade text filters. Drawing on prior demonstrations, the article distinguishes Unicode Tag payloads that some models interpret as readable instructions from zero-width binary encodings that generally require an explicit decoding scheme.