An AI Hacking Trick: Splitting ‘Funding’ in Spam
Microsoft has uncovered a unique and rather mundane use of a technique previously employed for more nefarious purposes.
Invisible Unicode characters, once used to hide instructions from both humans and AI models, have now been utilized in a phishing operation with a different aim. Instead of concealing commands, these characters are being employed to split the word "funding" in spam messages, aiming to bypass filters that match literal strings.
The ASCII smuggling campaign, detected during September 2026, reached an impressive 2.37 million messages per weekday. However, despite its sophistication, it ultimately failed, as over 99% of the messages were successfully intercepted by various security measures, including sender reputation checks, URL analysis, and machine learning classifiers.
The technique involves manipulating the Unicode Tags block (U+E0000 to U+E007F), which contains invisible shadows of printable ASCII characters. By inserting these characters into words like "funding," spammers hoped to create token fragmentation, making it harder for spam filters to identify and block their messages.
As Microsoft’s security researchers point out, this approach is a "game changer" in the world of phishing, demonstrating how even seemingly harmless Unicode characters can be exploited for malicious purposes.