A technique once used primarily to hide malicious prompts in AI agent attacks has found a new application with spammers seeking to evade email filters, according to a report.
What Happened
ASCII smuggling uses special Unicode tags to encode text readable by computers but nearly invisible to humans. The technique employs characters from block U+E0041 through U+E0061, which mirror standard ASCII letters like 'A' and 'a'. By embedding instructions in these hidden tags, large language models can detect the content while human readers see nothing unusual. The approach gained attention roughly two years ago as a method for making prompt injection attacks more stealthy. Spammers have now adopted the same technique to disguise mass email campaigns from filters designed to flag unwanted messages.
Why It Matters
The crossover illustrates how techniques developed in the AI security research community are being repurposed by malicious actors outside that domain. Email platforms rely on filtering systems that scan for known patterns associated with spam. ASCII smuggling allows those patterns to be hidden within characters that traditional detection tools may not process correctly, while LLMs used for content analysis can still parse them normally. For developers building AI-powered security tools and email platforms alike, the technique highlights an ongoing arms race between filter evasion methods and detection capabilities.
The Bottom Line
ASCII smuggling has evolved from a prompt injection defense into a spam evasion tool as bad actors adopt techniques originally explored in AI security research.