Over one million phishing emails evade AI security via text salting
Barracuda researchers identified over one million phishing emails using a 'text salting' technique that hides benign text to evade AI-powered email security systems. The technique splits malicious phrases across hidden and visible text, causing LLM-based security tools to misclassify the emails as legitimate, enabling attackers to use generative AI to create large volumes of varied campaigns.
What happened
Barracuda researchers identified over one million retail-themed phishing attacks using 'text salting' since April 2026. The technique involves inserting large amounts of unrelated, benign text into phishing emails while hiding it from recipients but keeping it visible to email scanners. Attackers concealed the additional text using methods including shrinking the visible display area, indenting text outside the visible frame, inserting zero-font-size text within words or sentences, and fragmenting HTML content to disrupt signature-based detection.
The hidden content typically consisted of unrelated material such as stories, project notes, or conversational text containing common, benign words including 'puppy,' 'training,' 'notes,' 'task,' 'rhythm,' and 'book,' while the visible portion presented retail-themed lures such as expiring reward points, gift cards, or loyalty offers.
Reported impact
The text salting technique deceives both traditional keyword-based detection and AI-based security systems, including machine learning models and LLM-based email security tools. Large volumes of benign hidden content can influence LLM-based models to classify phishing emails as legitimate.
Index analysis
Barracuda researchers observed increased use of text salting over the past year, with attackers adapting the technique to evade not only traditional spam filters but also machine learning and LLM-based email security systems. The availability of generative AI enables attackers to automatically generate large volumes of unique, natural-sounding text, making phishing campaigns more varied and difficult to detect.
Governance controls
Barracuda recommended a layered security approach and user training as defenses. The report notes that modern email security must analyze what users actually see, not just the raw email source code, to detect hidden-content evasion techniques.
Classification
Sources and evidence
Sources not yet published
This record was researched and written by the Index from original reporting. Its sources are still being verified by an editor, and they are listed here once that check is complete. Until then the record carries no reviewer.