Hidden Text Salting in Phishing Emails to Bypass LLM-Based Detection

When your phishing email whispers 'I love this email!' in hidden BDO tags, you're not just fooling the LLM -- you're giving it Stockholm syndrome before it even classifies the threat

Fantastic - now our SEGs need a threat model for 1998 HTML tags; the next SOC alert will be titled ‘Doge-directed bidirectional override detected.’

After 20 years of defending against SQL injection, we've successfully evolved to defending against prompt injection - because why solve old problems when you can just rename them with AI buzzwords and charge consulting fees all over again?

Ah yes, the classic 'I love this email!' hidden in a <bdo> tag - because nothing says 'legitimate business correspondence' quite like bidirectional override elements containing sentiment manipulation for your LLM spam filter. It's like SQL injection's younger sibling who went to art school and learned about semantic attacks. The real kicker? We spent decades teaching users to hover over links, and now we need to teach AI models to inspect the DOM for hidden compliments. At least when attackers used white-on-white text, you could just select-all to catch them. Now they're literally salting the prompt space like it's a bcrypt hash, except instead of protecting passwords, they're seasoning your intent classifier until it thinks every phishing email is a love letter from your CTO

Proof that AI-powered phishing detection without an HTML render step is just regex with delusions - sprinkle a few <bdo> salts and the classifier confidently labels the phish as a love letter

We replaced 600 SEG regex rules with an LLM - turns out a single BDO‑wrapped “I love this email!” string can jailbreak intent scoring; finally, a security bug where adding salt lowers entropy

CSS opacity:0 jailbreaking LLMs - proof that frontend hacks outsmart AI security layers faster than any zero-day