Microsoft uncovers Unicode‑tag spam that splits “funding” to evade filters
Microsoft Security Research identified a spam campaign that used invisible Unicode tag characters to split the word “funding” inside email bodies, allowing the messages to slip past keyword‑based filters. The operation, reported by researchers Noam Kochavi and Sarah Wolstencroft on Thursday, peaked at more than two million messages per weekday before collapsing in late May.
Key Takeaways
- The campaign injected invisible Unicode tag characters (U+E0000‑U+E007F) into the middle of words like “funding” to break tokenisation.
- More than 99 % of the spam was blocked by conventional defenses such as reputation checks, URL analysis and OCR, not by the Unicode‑tag detection.
- At its height the operation sent between 1 million and 2.37 million messages per weekday from roughly 150 disposable finance‑themed domains.
- ActiveCampaign, a legitimate marketing platform, was used as the relay, with about 92 % of the volume originating from a single network block belonging to the service.
- Microsoft’s new hunting signature now treats the presence of tag characters as a high‑confidence indicator of malicious email.
| Metric | Value | Why it matters |
|---|---|---|
| Technique name | ASCII smuggling (Unicode Tags block) | Enables invisible characters that can alter tokenisation without visible changes. |
| Peak daily volume | Between 2.3 million and 2.5 million messages | Shows the scale of the evasion attempt and the load on email defenses. |
| Date of first large spike | 8 February 2024 (≈21 000 alerts) → 9 February 2024 (>1.3 million alerts) | Marks the moment Microsoft’s signature began detecting the tag characters. |
| Number of sender domains | ≈150 disposable finance‑themed domains (148 on 9 Feb) | Illustrates the breadth of the infrastructure used for bulk sending. |
What is ASCII smuggling and why it matters to AI security
The Unicode Tags block (U+E0000‑U+E007F) was originally intended for language tagging and later considered for regional flag emojis, but it never saw mainstream adoption. Each code point mirrors a printable ASCII character—U+E0041 corresponds to “A”, U+E0061 to “a”—yet the characters remain invisible in standard text rendering. AI researchers have exploited this “shadow copy” to embed hidden prompts that large language models (LLMs) read and obey, a technique known as prompt injection.
Typical examples include inserting a hidden line in a Word document that rewrites spreadsheet values, or feeding a phishing email to an autonomous agent that then leaks AWS credentials. In every case the machine processes the invisible instruction while the human user sees nothing.
How the spam campaign repurposed the tags
Microsoft’s hunting signature scanned inbound email for any of the tag characters. When alerts surged, analysts examined a sample and discovered that the characters were not being used for hidden commands. Instead, they were placed inside ordinary words. By inserting a single invisible tag between “fun” and “ding”, the word “funding” appeared unchanged to the recipient, but token‑based spam classifiers saw two separate fragments that did not match known malicious keywords.
“The intent is inverted, but the mechanism is similar, and a user’s suspicions are not raised,” Microsoft wrote.
This subtle manipulation aimed to bypass keyword matching while keeping the visual content identical, exploiting the fact that modern machine‑learning classifiers first split text into tokens before applying any higher‑level reasoning.
Scale, timing and the infrastructure behind the spam
The campaign ran on a scheduled bulk‑sending infrastructure, delivering messages every weekday and dropping to near‑zero on Sundays. Daily volumes ranged from 1 million to a reported 2.37 million messages for roughly three months. By late March, weekday traffic had fallen about 80 %, and after 15 May the flow sharply declined, leaving only a low‑level residue into mid‑June.
Approximately 150 disposable domains, assembled from a 28‑word vocabulary (e.g., “advance”, “boost”, “finance”, “guardian”, “wave”), were used as sender addresses. On 9 February alone, 148 such domains were observed, generating up to 30 442 hits for the busiest domain in a single day.
The actual sending was performed through ActiveCampaign, a legitimate marketing‑automation platform. ActiveCampaign rewrites every link to route through its own tracking domains, and about 92 % of the spam volume originated from a single network block owned by the service. The platform confirmed that its moderation system treated the obfuscated messages the same as clean ones, though it flags heavy use of invisible characters as suspicious.
Why the evasion failed and what defenses caught the bulk
Despite the clever token‑splitting, more than 99 % of the messages were stopped by layers unrelated to the Unicode tags. Sender and IP reputation checks, URL and domain reputation, brand‑impersonation detection, authentication failures, and the primary machine‑learning classifiers all triggered. Additionally, Microsoft Defender’s OCR step re‑reads the email content as an image, reproducing exactly what a human sees and thereby neutralising the invisible characters.
Because tag characters are exceedingly rare in normal email traffic, their presence became a high‑confidence indicator of malicious intent. The initial hunting signature even produced a false positive on legitimate emails containing the flag emojis of England, Scotland and Wales—emoji sequences that are constructed from the same invisible tag characters.
Microsoft’s response and recommendations
Following the discovery, Microsoft released a signature that flags any occurrence of Unicode tag characters in email bodies. The signature’s first iteration mistakenly flagged legitimate flag‑emoji messages, prompting a quick refinement to exclude those cases. Microsoft advises security teams to normalise incoming text—stripping or collapsing invisible tag characters—before performing token‑based analysis or keyword matching.
Why it matters
The episode demonstrates how a technique originally devised for AI prompt injection can be repurposed for conventional spam, highlighting the fluid boundary between emerging AI threats and traditional email abuse. It also underscores the importance of layered defenses: even sophisticated evasion tactics can be mitigated when reputation, URL analysis, OCR and behavioural checks are combined. Finally, the reliance on a reputable marketing platform illustrates how attackers increasingly piggyback on trusted infrastructure to amplify reach, a trend that security teams must monitor closely.
Frequently asked questions
Who discovered the Unicode‑tag spam campaign?
The campaign was uncovered by Noam Kochavi and Sarah Wolstencroft of Microsoft Security Research, who detailed their findings in a blog post on Thursday.
What Unicode block was exploited?
Attackers used the Unicode Tags block (U+E0000‑U+E007F), which provides invisible shadow copies of the standard ASCII characters.
How many messages were sent during the peak of the operation?
Peak daily volumes were reported between 2.3 million and 2.5 million messages, with weekday traffic ranging from 1 million to 2.37 million messages for about three months.
Why didn’t the invisible characters alone bypass Microsoft’s defenses?
More than 99 % of the spam was blocked by conventional layers such as reputation checks, URL analysis, brand‑impersonation detection, authentication failures, and OCR‑based content verification.
What platform did the attackers use to relay the spam?
The bulk of the messages were sent through ActiveCampaign, a legitimate marketing automation service, with roughly 92 % of the volume originating from a single network block owned by that platform.
Bottom Line
Microsoft’s investigation reveals that invisible Unicode tag characters can be weaponised to split words and evade token‑based spam filters, but the technique alone is insufficient against a robust, multi‑layered email security stack. The campaign’s reliance on a reputable marketing platform further illustrates the evolving tactics of threat actors, reinforcing the need for continuous monitoring and adaptive defenses.
As email ecosystems continue to integrate AI‑driven components, security teams should normalise text inputs and remain vigilant for rare Unicode patterns that may signal novel evasion attempts.
Reporting based on: original source.
Comments
Post a Comment