Homoglyph Attack
A homoglyph attack substitutes characters that look identical or nearly so but are different code points: a Cyrillic а for a Latin a, a Greek ο for a Latin o, a digit 1 for a lowercase l. In a domain name it produces a look-alike address for phishing; in text fed to a language model it disguises a keyword so that a filter misses it while the model still reads it.
Unicode contains many characters that render alike. The classic use is the internationalised domain name that displays as a trusted brand but resolves somewhere else; browsers now show such names in their raw punycode form when the script mix is suspicious, which blunted but did not end the technique. Registrars, e-mail clients and chat applications remain uneven in how they display mixed-script strings.
Against AI systems the target is the filter rather than the human. A scanner looking for “ignore previous instructions” does not match “ignоre” with a Cyrillic о, but a language model, trained on noisy text, reads the word as intended. Homoglyphs, like invisible characters and encoding, are therefore an evasion layer wrapped around another injection technique.
The defence is normalisation. Fold the text to a canonical form (NFKC), map confusable characters to their Latin equivalents, and count how many scripts are mixed within a single word; then run detection on the normalised text while keeping a map back to the original positions so the flagged span is reported correctly.
Example
A phishing e-mail links to a domain that renders as a well-known payment brand. Inspected, the second letter is U+0430 (Cyrillic small a) rather than U+0061; the domain was registered two days earlier and resolves to a hosting address listed for phishing. The same substitution, applied to the words of an instruction inside a web page, slips past a naive keyword filter.
In isMalicious
The gate normalises every scanned document (NFKC, confusable folding, script mixing) before it runs the detection families, and homoglyph is itself one of those families, so a mixed-script word is flagged and the underlying instruction is detected in its folded form. A domain report shows the punycode form of an internationalised name alongside its registration age.
Frequently Asked Questions
What is Homoglyph Attack?
A homoglyph attack substitutes characters that look identical or nearly so but are different code points: a Cyrillic а for a Latin a, a Greek ο for a Latin o, a digit 1 for a lowercase l. In a domain name it produces a look-alike address for phishing; in text fed to a language model it disguises a keyword so that a filter misses it while the model still reads it.
How is Homoglyph Attack related to Typosquatting?
Homoglyph Attack and Typosquatting are both key concepts in threat intelligence. Typosquatting (also called URL hijacking) registers domains that are slight misspellings of legitimate websites to capture traffic from users who make typing errors. These domains are often used for phishing, malware distribution, or ad fraud.
Related Terms
Typosquatting
Typosquatting (also called URL hijacking) registers domains that are slight misspellings of legitimate websites to capture traffic from users who make typing errors. These domains are often used for phishing, malware distribution, or ad fraud.
Invisible Text Injection
Invisible text injection hides instructions for a language model in content a human cannot see: zero-width characters, Unicode tag characters, text coloured to match its background, font size zero, HTML comments, off-screen elements, or alt text and metadata. The reviewer sees a normal document; the model, which reads the raw text, sees the instructions.
Phishing
Phishing is a social engineering attack that tricks users into revealing credentials, clicking malicious links, or downloading malware — typically via email. Spear phishing targets specific individuals; smishing uses SMS; vishing uses voice calls.
Put this intelligence to work
Query indexed indicators — IPs, domains, URLs, and hashes — in seconds.