One technique is simultaneously a weapon aimed at large language models and a workhorse for two-bit email spammers. ASCII smuggling started this decade as a way to slip hidden instructions past an AI’s safety filters. In February 2026, Microsoft watched the same trick migrate into ordinary phishing emails, with a sharp rise in detections. The tool didn’t change. The target did.
Technically speaking, ASCII smuggling is a payload built from invisible Unicode characters — tag characters like U+2060 (word joiner) and directional marks that render as nothing to the human eye but persist fully intact at the text-processing layer. Spammers interleave these characters between the letters of a malicious string. The email’s recipient sees a harmless-looking URL. The filter, if it inspects the raw text, sees a scrambled sequence that defeats its signature-based regex patterns and its link-extraction heuristics. On the AI side, the same invisible characters can be used to hide a system-breaking instruction from the model’s alignment trainare who they’re scampering past. The only thing that changed is who got caught.
🕒 Published: