Once popular for attacking AI, ASCII smuggling is embraced by spammers
A once-overlooked block of unicode that's invisible to humans is gaining ever wider use.
ASCII smuggling, a technique once favored for concealing malicious prompts in attacks against AI, is now being utilized by spammers to bypass email filters. This method, known as ASCII smuggling, surfaced two years ago as a way to make AI attacks, specifically prompt injections, less conspicuous. These attacks are carried out by embedding instructions in emails or other untrusted content intended for processing by large language models (LLMs).
However, these instructions are not written in standard text; they are encoded using a special set of Unicode tags. For instance, the Unicode point U+E0041 resembles the letter "A," while U+E0061 mimics the lowercase "a." These tags closely resemble characters from the American Standard Code for Information Interchange (ASCII), but with a key distinction: they are readable by computers but virtually invisible to humans.
By encoding malicious prompts in these tags, LLMs can detect the instructions, while people reading the email are none the wiser. For further details on ASCII smuggling, refer to the linked article.
Written by urgent.news from Ars Technica's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
