- Invisible-character obfuscation is a long-standing email evasion method. The Unicode Tags block became widely known through AI prompt-injection research, but the property that makes it useful there is text a person can’t see and software still processes. That same property lets attackers break words apart so detectors miss them while readers don’t.
- Mimecast observed the technique across bulk finance phishing, compromised-account credential lures, call to action URL and brand spoofing, including mid-word splices, saturated subjects, and homoglyph display names-
Hidden in plain sight: How invisible Unicode helps evade filters
By Mimecast Threat Research Team
Key Points
Threat Overview
Obfuscated text is as old as text processing itself, and attackers have cycled through homoglyphs, character padding, image-only bodies and HTML fragmentation for decades. What has changed is the vocabulary. Over recent months Mimecast Threat Research continues to observe the use of various obfuscation techniques including the use of invisible Unicode characters. Code points that occupy a position in the raw byte stream but render as nothing in a normal mail client placed inside lure keywords.
The effect is a deliberate divergence between two readers of the same message. The recipient sees the words. A detection engine performing a literal string match sees something different. No vulnerability is exploited; no payload is delivered by the technique itself, and the cost to the operator is a single find-and-replace in a template.
What It Is
The main vehicle is the Unicode Tags block (U+E0000–U+E007F). Each code point sits at a fixed offset from a printable ASCII character (U+E0020 for a space, U+E0043 for ”C”), but typical fonts and mail clients render none of them. The block was designed for language tagging, a use that has since been abandoned.
Because every ASCII character has an invisible twin, an attacker can write instructions that a human reviewer never sees but a language model reads as ordinary text. That is why the block features so heavily in prompt-injection research.
In email evasion the goal is different. The attacker isn’t hiding a message. They place an invisible character in the middle of a keyword so that contiguous string matches no longer fire.
Why It Works
Many spam and phishing controls still rest on literal keyword lists or regular expressions. Interleaving an invisible character mid-word defeats those checks without changing a single visible glyph: Three properties make this attractive to an operator:
- No vulnerability is required. The behavior is correct Unicode handling by every component in the path. Nothing is malformed, so no errors.
- The cost is negligible. A template change applies the technique to an entire campaign, and the character can be moved to a different position in the word on every send to frustrate signature writing.
- It degrades machine-learning classifiers too. Without normalization, a tokenizer that meets an unfamiliar code point may break a telling term like ”credit” into meaningless fragments. That weakens ML classifiers as well as keyword rules.
How It Is Encoded
The same code point can reach a mailbox in several representations, and detection content must account for all of them. Using U+E0020 (TAG SPACE) as in Table 1.
| Representation | Form | Where observed |
|---|---|---|
| Unicode code point | U+E0020 | Notation only |
| Raw UTF-8 | F3 A0 80 A0 | HTML body part, base64/quoted-printable encoded |
| HTML decimal entity | | Plain-text alternative body part |
| HTML hex entity | | HTML body part |
| MIME encoded-word | =?UTF-8?B?…?= / =?UTF-8?Q?…?= | Subject and From headers |
Table 1: Representation of U+E0020 observed in email
Observed Campaigns
Campaign One: Minimal Splicing in Bulk Funding Lures in Body of Email.
The most widespread usage Mimecast has observed is also the most economical. Messages advertise unsolicited business funding, credit lines and advance-funding products, and drop a single TAG SPACE inside each finance keyword rather than encoding any hidden message.
In one campaign, seven keywords were fractured with U+E0020 across a 14,283-character HTML body an obfuscation footprint of well under 0.1%, targeted precisely at the terms a keyword rule would look for:
| Raw string | Rendered to recipient |
|---|---|
| Cre + U+E0020 + dit | Credit |
| ter + U+E0020 + ms | terms |
| ra + U+E0020 + tes | rates |
| acc + U+E0020 + ess | access |
| Ac + U+E0020 + cess | Access |
| fa + U+E0020 + st | fast |
Table 2: U+E0020 keyword splits and their rendered forms
Figure 1: Phishing campaign fractured with U+E0020
Note the selection. Every fractured term is a funding-vocabulary keyword or an urgency modifier. Surrounding texts such as Bridge Funding, prepayment fees, and the sender's job title were left untouched, indicating a curated substitution list rather than blanket randomization.
Campaign Two: Saturation Obfuscation in Subject from a Compromised Account
A second and considerably more aggressive pattern was observed in messages sent from a compromised education-sector mailbox in June and July 2026. Here the operator did not splice selected keywords; they saturated the entire subject line.
| Metric | Account-expiry lure (Jun 2026) | Payslip lure (Jul 2026) |
|---|---|---|
| Subject as rendered | University Email Account Access Ending Soon 372508BQB | Your Payslip Has Been Generated 361432LWA |
| Visible characters | 53 | 41 |
| Invisible code points | 324 | 261 |
| Proportion invisible | 85.9% | 86.4% |
| Distinct invisible characters | 34 | 34 |
| Invisible-to-visible ratio | 6.1 : 1 | 6.4 : 1 |
Table 3: Invisible character density in saturated subject lines
Figure 2: Phishing Campaign using invisible characters
Every boundary between visible characters carries a randomized run of six to eight invisible code points mixed across all three families. A single boundary from the June sample illustrates the construction:
| 'S' + [U+00AD, U+2063, U+FEFF, U+2060, U+00AD, U+FE00] + 'o' -> "So" 'n' + [U+2063, U+2064, U+00AD, U+2061, U+2061, U+FEFF, U+2062, U+FE0B] + 'g' -> "ng" |
This defeats substring matching at any length greater than one character and, critically, also breaks the alphanumeric campaign identifiers appended to the subject (372508BQB, 361432LWA) that would otherwise cluster the campaign trivially.
The Tags characters present in these subjects do not spell a coherent hidden message. Projecting them to ASCII yields punctuation noise such as +&**#( '''$!)), confirming that the objective is fragmentation rather than smuggled content — a meaningful distinction from the AI prompt-injection use case that shares the same primitive.
Figure 3: Phishing campaign within minimal content in body
In the plain-text body of the same university-access sample, obfuscation was not limited to the subject. Immediately after the greeting, the MIME used quoted printable to insert invisible code points that the recipient still reads as a normal ”Hello Corinne,” line:
| Hell=E2=80=8Co Corinne, =C2=AD=E2=80=8E=E2=80=8E=E2=80=8D=E2=80=8F=E2=81=A1=C2=AD=E2=81=A3=E2=80=8F= =E2=81=A0=EF=B8=87=EF=B8=83=EF=B8=80=EF=B8=85=F3=A0=80=A5=F3=A0=80=A7=F3=A0= =80=A6 --=20 |
Figure 4: Quoted-printable greeting and invisible code points
Decoded, that becomes:
- Hell + U+200C (ZERO WIDTH NON-JOINER; QP =E2=80=8C) + o Corinne, → user sees Hello Corinne,
- Then a compact invisible run before the signature separator --:
So the body carries both a mid-word ZWNJ splice in Hello and a Tags-block-inclusive invisible cluster, the same toolkit seen in the fractured subject while remaining visually benign in a normal mail client.
Payload
Each message carried a Word document using account-expiry document lure with a Microsoft "Review or Update Your Microsoft Account" notice claiming the recipient's Office 365 mailbox was scheduled for termination. The payslip document reproduced an employee payroll-portal notification and linked to a Cloudflare Workers subdomain.
Campaign Three: Homoglyph Characters in the Sender Display Name of MIME Header
A third variant applies obfuscation to sender identity rather than body content. In an August 2026 banking-alert sample, the display name was encoded as a MIME encoded-word:
| From: =?UTF-8?Q?B=C9=91nk_of_America?= <ealerts.backofamerica@pedro-tuning[.]sk> |
The quoted-printable sequence =C9=91 decodes to U+0251 LATIN SMALL LETTER ALPHA, which is visually near-identical to a lowercase a in most sans-serif fonts. The display name renders as Bank of America while never containing the literal string Bank of America that a brand-impersonation rule would match. Unlike the invisible-character technique, this substitution is visible but only to a reader who is looking for it.
Figure 5: Phishing campaign using ”Bank of America” display name containing U+0251 (Latin small letter alpha)
Campaign Four: Call Back Phishing with zero-width space encoding in Brand Impersonation.
Amazon ”order confirmation / payment success” subject carrying a full scam narrative and callback number, delivered via Amazon SES branding.
Figure 6: Amazon phishing sample using invisible character fracture in brand text
Classic invisible-character fracture in this family of campaigns often uses zero-width space (ZWSP):
Example: Ama + U+200B + zon atill reads Amazon to the user, but literal keyword/signature matches the contiguous string Amazon can fail unless text is normalized first.
This sits in the same obfuscation family as Tags-block splices and dense subject fracturing: the recipient sees intact brand/lure language; detectors that match unbroken tokens may not.
Campaign Five: Credential Phishing with Zero-Width Character in CTA (Call to Action)
As shown in figure 7.0, the message is addressed generically to ”Dear Team,” signed ”Payroll/Human Resources,” and states that the attached ”Layoff_Letter.pdf” lists the affected staff. The href points to https://google.com.tw/share.google?q=<token>#<recipient_email>.
Figure 7.0: Phishing Email having Zero width character in CTA
In the raw HTML, as showin figure 8.0, the hostname is written with three numeric character references to U+180E (). A browser decodes these into invisible characters, so the status bar and hover preview show what looks like google.com.tw, while the underlying string differs. This deceives the recipient, and it can also confuse tools that decode and strip format characters before applying an allowlist, or that compare the host with exact string matching.
Figure 8.0: Zero Width character in CTA
The URL fragment carries the recipient’s email address (#user@domain). Fragments are not transmitted in HTTP requests, so the value does not appear in server or proxy logs, but client-side JavaScript on the landing page can read it from location.hash, which phishing kits commonly use to pre-fill a credential form. The message is also delivered as quoted-printable (=3D for =, soft line breaks), so parsers must decode that layer before they can see the URL. Reliable detection therefore requires a canonicalization pipeline: decode quoted-printable, unescape HTML entities, strip or flag format characters (U+180E, U+200B to U+200D, U+2060, U+FEFF), then parse the URL and evaluate both the raw and normalized forms.
Risk in Both Directions
This technique is not new. Invisible and homoglyph obfuscation has circulated in spam tooling for many years, and the Tags block has been a known curiosity for longer. What has changed is the surrounding conversation.
Prompt-injection research made ASCII smuggling a widely documented, well-tooled and well-understood primitive. Public encoders and decoders now exist; proof-of-concept payloads are trivial to find, and the technique has an established name. That research raised the profile of the primitive far beyond its previous audience, and the same encoders that produce hidden model instructions produce hidden filter breaks.
The direction of concern now runs both ways, and organizations should plan for both:
- Invisible characters as an evasion. The usage documented above, aimed at keyword rules, signatures, and classifier tokenizers.
- Invisible characters as instructions. As AI assistants and copilots are given access to mailboxes to summarize, triage, draft replies and take action, a message body becomes an untrusted input to a model. Content that is invisible to the human approving the action is not invisible to the assistant reading it.
Detection Observations
Content-only evaluation is the weakest position against this technique. Effective detection combines the message body with sending patterns, age of indicators, anomalous sender behavior and intent analysis, so a technique that only fractures keywords are less likely to succeed on string-matching gaps.
Mimecast’s Multi Vector Threat Protection (MVTP) combines the message body with the structure of the email, age of indicators, and anomalous sender behavior — among other signals derived from behavioral AI to reach a verdict, making it resilient to techniques that fracture keywords. Besides, MVTP, Mimecast's AI classifiers as well are not defeated by the invisible-character splicing. Before sub word tokenization, the text-cleaning stage removes Unicode format (Cf) code points — covering the Tags block (U+E0000–U+E007F), zero-width and bidirectional format characters (U+200B–U+200F, U+2060–U+2064), the soft hyphen (U+00AD) and the byte-order mark (U+FEFF). A word spliced with characters from any of these ranges produces the same token sequence as the clean word: Cre<U+E0020>dit and Credit are indistinguishable to the model. The classifier sees the text as the recipient reads it, which is why keyword fracturing that defeats a literal string match does not carry over to model-based detection.
Figure 7: Detection across seven months
As shown in figure 7.0, detection volume spiked sharply in early March, peaking near 26,000 in a single day, then swung between roughly 10,000 and 22,000 through mid-April — consistent with attackers adjusting their approach to evade detection. From late April onward, volume dropped steadily, settling into a low, stable baseline through September.
While this decline is consistent with our defenses closing the gap, detection volume alone doesn't confirm whether attackers abandoned the technique or simply shifted elsewhere. We're continuing to monitor for resurgence and validate this trend against external threat intelligence.
Conclusion
Obfuscated text is as old as text processing itself. Mimecast Threat Research has observed a renewed increase in invisible Tags-block characters used for that purpose. The technique, commonly known as ASCII smuggling and popularized by AI prompt-injection research, is still used to evade email keyword and signature detection: attackers insert invisible characters inside words so recipients see normal text while literal filters often do not.
Mimecast has seen the technique across bulk finance phishing, compromised account credential lures, brand spoofing, and callback scams. Earlier campaigns were sent in volume through a mainstream marketing platform from disposable, finance-themed domains, with the familiar aim of unsolicited loan, credit-line, and advance-funding phishing tied to fraud and credential theft. The same method later appeared in messages from compromised accounts, including business email compromise and callback themes, as well as homoglyph display names and heavily fractured subjects. Daily volume surged in March, peaking near 26,000, then fell sharply through April and stayed low and sporadic from May into September.
Mimecast’s Multi-Vector Threat Protection reaches a verdict from the message body together with email structure, indicator age, and anomalous sender behavior. Mimecast’s AI classifiers also strip Unicode format characters, including the Tags block, before tokenization, so a spliced word such as ”Credit” produces the same token sequence as the clean word. Keyword fracturing that defeats a literal string match does not carry over to that model-based detection.
Indicators of Compromise
Campaign One — bulk funding lure
| Indicator | Type |
|---|---|
| expressloccapital[.]com | Domain |
| em-3251641.expressloccapital[.]com | Domain |
| expressloccapital123.acemlnd[.]com | Domain |
| expressloccapital123.activehosted[.]com | Domain |
| [External] → Quick note on <company> funding options | Subject |
Campaign Two — compromised-account credential phishing
| Indicator | Type |
|---|---|
| msoff.systematicapproach[.]one | Domain |
| solutionalliancetacticblue-waterfall-e044.mihofo9840.workers[.]dev | Domain |
| Secure_Document_Copy_55338.docx | Filename |
| fa540148c96fa80cd1d9591f73c4ed4e3a66679cc7819427efc386184c6fd017 | SHA-256 |
| Compliance_Record_Document_67546.docx | Filename |
| 4292da028ac1fac39eed78b4040086ddf53226d022658fb4d7b8c121fead6e5c | SHA-256 |
| University Email Account Access Ending Soon 372508BQB | Subject |
| Your Payslip Has Been Generated 361432LWA | Subject |
| Mailbird 2.9.80.0 / Postfix MTA 3.6.7 | X-Mailer |
Campaign Three — homoglyph banking alert
| Indicator | Type | Notes |
|---|---|---|
| ealerts.backofamerica@pedro-tuning[.]sk | Sender | Display name Bɑnk of America using U+0251 |
| links.notification.intuit[.]com | Domain | Legitimate click-tracking service abused for redirection |