Researchers at Barracuda Networks uncovered more than a million retail‑focused phishing emails that use hidden text, known as text salting, to slip past AI‑driven security gateways. The technique exploits weaknesses in large language models, allowing attackers to craft convincing messages while evading detection.

Key Takeaways

  • Hidden text (text salting) lets over 1 million phishing emails evade AI filters.
  • Large language models accelerate the creation of salted content, outpacing current defenses.
  • Effective protection requires context‑aware analysis beyond keyword detection.

From April onward, researchers at Barracuda Networks observed a wave of retail‑themed phishing campaigns that exceeded one million messages. These emails embed hidden text – a technique called “text salting” – to confuse both static and AI‑powered email security gateways, rendering traditional keyword‑based detection ineffective.

What Is Text Salting?

Text salting is an old‑school method where attackers insert benign or nonsensical words into the HTML of an email and then hide them from view. Techniques include setting the font size to zero, wrapping filler text in containers with zero height and width, or pushing the content off‑screen so no scroll bar reveals it. To a human reader, the visible portion appears legitimate, while the hidden filler dilutes spam‑my patterns that filters rely on.

LLMs Give the Trick New Speed

Large language models (LLMs) have turned this manual process into an automated one. By prompting an LLM, attackers can generate vast amounts of innocuous filler text—random stories, harmless words like “puppy” or “book”—and embed them instantly. Meanwhile, AI‑based security engines, which often depend on keyword matching and simple heuristics, struggle to parse the layered, multi‑technique salting that now overwhelms them.

Why Secure Email Gateways Miss These Emails

Secure Email Gateways (SEGs) perform two main checks: authentication (DKIM, SPF, domain reputation) and content analysis. Most of the observed campaigns passed authentication, using either legitimate domains or compromised sites that mimic brand names. The real challenge lies in content analysis. While the visible text contains typical phishing cues—“urgent,” “rewards,” “Congratulations!”—the hidden filler breaks up these cues, preventing the engine from flagging the message.

Moving Toward Context‑Aware Defense

Experts argue that defending against text salting requires more than scanning for suspicious keywords. Security tools must evaluate the full context: the relationship between visible content, hidden or excessive text, link behavior, sender reputation, and the intended user action. Modern defenses are beginning to un‑hide invisible text by parsing CSS and HTML rules, but attackers counteract by layering multiple salting techniques, ensuring that if one method is uncovered, another remains concealed.

As LLMs continue to lower the cost of generating sophisticated phishing content, the prevalence of text‑salting attacks is likely to rise. Organizations should therefore revisit their email security architectures, integrating deeper content inspection, behavior‑based analytics, and continuous AI model updates to stay ahead of this evolving threat.