Modern email security is facing a nostalgic challenge as researchers discover that classic 'text salting' techniques—tactics used decades ago to fool basic filters—are successfully bypassing sophisticated Large Language Model (LLM) spam detectors.
The Return of Text Salting
Text salting involves inserting invisible or irrelevant strings of characters into an email to dilute the presence of 'spammy' keywords. While these tricks were largely neutralized by traditional Bayesian filters, the current generation of AI-powered filters appears susceptible to the same manipulation. By adding innocuous text or 'noise,' attackers can confuse the LLM's context window, leading the system to misclassify malicious content as legitimate correspondence.
The resurgence of these old-school methods highlights a potential blind spot in AI-driven security: the tendency for models to be overly influenced by the overall distribution of words rather than identifying specific patterns of intent. As attackers revert to these simple yet effective methods, security providers may need to rethink how LLMs process email metadata and hidden content.
