ASCII smuggling is a technique that hides text inside ordinary-looking content using invisible Unicode characters, usually from the Unicode Tags block (U+E0000 to U+E007F). Readers see nothing unusual. But any software that processes raw characters like artificial intelligence (AI) chatbots and email filters reads the hidden message or gets tricked by it.
Think of it as a note written in invisible ink between the lines of a normal letter. Your coworker skims it and sees a friendly lunch invite. Someone holding an ultraviolet (UV) flashlight sees a second message meant only for him/her. In the technique, AI models and text-processing tools are like the people holding such a flashlight.
Table of Contents
- How the Technique Works
- Why Spammers Embraced the Technique
- The Technique Amid the Proliferation of AI
- Sample Vulnerabilities
- Attack Types
- Attack Vectors
- Attack Implications
- Best Practices and Mitigations
- Is ASCII Smuggling Just a Security Risk?
Read More about ASCII Smuggling
The technique plants invisible Unicode characters in emails, documents, web pages, and prompts. AI models can read that hidden text and may follow it as instructions, while people reviewing the same content see nothing. Spammers also use it to split words so keyword filters miss them. The main fix is to strip or normalize invisible characters before any filter or model processes the text.
How the Technique Works
Two standards explain the trick. ASCII is the 1960s-era set of 128 characters covering English letters, digits, and basic punctuation. Unicode is its far larger successor, with more than 150,000 characters, including emoji and many special-purpose symbols.
One of those special-purpose ranges is the Tags block. Its 128 code points mirror ASCII one for one. The capital letter A is U+0041 and its tag twin is U+E0041. Tags were first meant to label a passage’s language. Today, they mainly build subdivision flag emojis like England’s and Scotland’s. The Unicode standard says software that doesn’t support tags should simply not display them.
That invisibility enables attacks. Attackers can shift each character in a sentence into its tag equivalent and paste it next to normal text. Browsers and email clients show only the normal text. A large language model (LLM), however, breaks the full input into tokens, tags included. Because tag characters appeared in training data, many models decode them into readable instructions.
What You See versus What a Machine Reads
A researcher showed how wide the gap between the two can be. While his Gemini prompt appeared to ask for five random words, hidden tags told it to print one specific word instead. Gemini obeyed the hidden text.
The technique follows four steps—encode, embed, pass review, and inject and act.
Why Spammers Embraced the Technique
Email filters often score messages by the words they contain. A note stuffed with “funding,” “invoice,” or “wire transfer” looks suspicious. Invisible characters let spammers break those words apart for the filter while keeping them intact for the reader.
Microsoft documented exactly this in September 2026. Researchers traced a phishing campaign that inserted an invisible tag space (U+E0020) inside lure words, so “funding” became “fun,” a hidden character, then “ding.” Keyword matching fails, yet recipients read the word normally. The emails came from about 150 finance-themed domains, mostly on weekdays.
The economics are hard to beat—no malware, no visual clue for victims, and a few lines of code to automate.
The Technique Amid the Proliferation of AI
Invisible-character tricks predate chatbots. The 2021 Trojan Source research made code look different to reviewers than to compilers. What changed was the arrival of AI systems that read every character.
In January 2024, researcher Riley Goodside showed ChatGPT following instructions hidden in tag characters. Johann Rehberger then released the ASCII Smuggler tool, which encodes and decodes such payloads.
The stakes rose as assistants gained access to inboxes, calendars, and tools. A hidden sentence in a chat prompt is a curiosity. The same sentence in an email read by an AI agent that can send messages is an attack path since hidden prompts can slip past people.
It also weakens “human in the loop” review, where a person approves AI actions. Approving what you can’t see means little.
Sample Vulnerabilities
Over time, we have seen several vulnerability exploitation attacks take advantage of the technique. Here are some of them.
1. Microsoft 365 Copilot Data Theft (2024)
A malicious email or shared document told Copilot to search the victim’s mailbox and files for sensitive data, which in the researcher’s demo included a Slack MFA code. Copilot hid the stolen data in invisible tag characters inside a link, and one click sent it to the attacker. It was reported to Microsoft in January 2024 and published that August, after Microsoft restricted how Copilot displays links.
2. From AI Prompt Injection to Phishing Evasion (2026)
In September 2026, Microsoft reported the reverse use. A hunt built to catch hidden prompts in email instead surfaced finance-themed phishing that spliced tag spaces into lure words to dodge filters. Hits jumped from about 21,000 on 8 February 2026 to more than 1.3 million the next day, peaking at 2.4 million on 26 February. Microsoft says its layered defenses flagged over 99% of the messages without relying on the tag characters.
Attack Types
Researchers usually group the technique by the invisible characters that carry the payload. Here’s a breakdown of its types.
| Type | How It Hides Data | Typical Use |
| Unicode Tags | Maps each ASCII character to its invisible twin in the U+E0000 to U+E007F range | Hidden prompts, keyword splitting |
| Variation selectors | Uses 256 invisible selectors (VS-1 to VS-256), each standing for one byte of data | Smuggling arbitrary data through copy-paste |
| Sneaky Bits | Encodes binary with two invisible characters like U+2062 for 0 and U+2064 for 1 | Data exfiltration |
| Zero-width characters | Inserts zero-width spaces and joiners (U+200B to U+200D) between visible letters | Filter evasion, text watermarking |
Tags remain most popular because the mapping is simple and many tokenizers decode it directly.
Attack Vectors
Hidden characters can ride along wherever text flows into a model or filter, such as in:
- Emails: Hidden instructions target AI assistants; split keywords target spam filters.
- Calendar invites: An invite can change what an assistant says about a meeting.
- Documents and PDFs: Files pulled into retrieval-augmented generation (RAG), where a model looks up documents before answering, can carry instructions.
- Web pages and forums: Text copied from a site or read by a browsing agent may hold a payload.
- Reviews and comments: A hidden link in a product review can surface in an AI-written review summary.
- Shared prompts and templates: Prompt libraries can include instructions the user never sees.
Attack Implications
The core problem is a split between what people approve and what machines execute. Identity spoofing results when an assistant repeats a fake organizer or sender. Data poisoning follows as summaries and reports quietly absorb attacker content, a concern for marketing teams whose reviews reach customers through AI summaries.
Assistants can also be told to tuck sensitive details into invisible text that leaves when someone copies or clicks. Filter evasion undermines exact-match controls, from spam rules to data loss prevention (DLP). Investigations suffer, too, since screenshots show only clean text.
Best Practices and Mitigations
As the technique gains popularity among attackers, security experts have identified mitigation strategies to address related threats. Here are some of them.
- Strip or normalize invisible code points first. Microsoft advises removing invisible characters before keyword or regex checks run. Researchers also recommend filtering tags in prompts and responses.
- Sanitize until nothing changes. Amazon Web Services (AWS) found single-pass Java filters can recombine stray surrogate halves into new tag characters, so it repeats cleanup and removes orphaned surrogates.
- Use platform guardrails. Amazon Bedrock Guardrails can block tag content and Microsoft Defender for AI services flags attacks.
- Monitor the raw input. Log the exact strings a model receives and set up alerts for tag sequences.
- Red-team your apps. Some solutions offer ASCII smuggling tests.
- Watch behavior. Microsoft treats weekday-only bursts of tag characters from churning finance domains as a strong signal of an attack.
- Train users. Experts advise verifying unexpected messages and avoiding unverified links.
Stripping tags turns flags like England’s into a plain black flag, and AWS notes attackers keep testing workarounds, so filters belong inside wider prompt-injection defenses. Read up on more ways to layer controls in this post.
Is ASCII Smuggling Just a Security Risk?
Not entirely, and some misconceptions need clearing up. The name misleads—nothing in plain ASCII is invisible; the technique uses Unicode characters that mirror ASCII. It isn’t only an AI problem either. The 2026 phishing emails carried no AI instructions, and one researcher warned early that the trick hides data in any text channel.
The characters also have honest uses. Tags make regional flags work, zero-width joiners combine emoji, and zero-width marks can watermark documents to trace leaks. Red teams plant smuggled payloads deliberately. And exposure varies by model, integration, and version.
Conclusion
The technique works because people and machines read the same text differently. Invisible Unicode tag characters let attackers hide instructions for AI assistants or split words to slip past spam filters, without leaving any visible trace. Real, documented cases show the technique moving from research into everyday attacks. Defense starts with a simple habit—strip or normalize invisible characters before any filter or model sees the text. Pair that with raw-input monitoring, red-team testing, and user awareness, and hidden text loses most of its power to fool people and machines alike.
FAQs
1. What is ASCII smuggling in simple terms?
It is a technique that hides a message inside normal text with invisible Unicode characters. People cannot see the message, but AI models and some software can read it or get confused by it.
2. Is the technique the same as prompt injection?
No. Prompt injection is the goal—sneaking instructions into an AI model’s input. ASCII smuggling is one way to deliver those instructions invisibly. It also works without AI as when dodging spam filters.
3. Which characters does the technique use?
Mostly Unicode tag characters (U+E0000 to U+E007F). Attackers also use zero-width spaces and joiners, variation selectors, and invisible math operators like U+2062 and U+2064.
4. How can I detect attacks?
Paste suspicious text into a decoder or scan it for code points in the ranges above. Organizations should log raw model inputs and set up alerts for those characters.
5. Can the technique affect regular email users?
Yes. Microsoft’s 2026 campaign used hidden characters to sneak phishing emails past filters. Treat unexpected financial messages with suspicion and confirm them through a separate channel.
Key Takeaways
- ASCII smuggling hides text with invisible Unicode characters, mainly from the Tags block, so people see clean content while machines read or misread hidden data.
- LLMs can decode tag characters because they appeared in training data, which lets attackers plant instructions that slip past human review and approval.
- Spammers adopted the technique in 2026, splitting financial keywords with invisible tag spaces to beat filters in a campaign that peaked at around 2.4 million messages.
- The strongest defense is stripping or normalizing invisible code points before filters and models process text, backed by raw-input monitoring, guardrails, and regular red-team testing.
- Tag characters also have legitimate uses, such as regional flag emojis, so blanket removal needs careful testing, and protection should sit inside broader prompt-injection defenses.
Sources
- https://aws.amazon.com/blogs/security/defending-llm-applications-against-unicode-character-smuggling/
- https://indianexpress.com/article/technology/tech-news-technology/what-is-ascii-smuggling-spam-ai-email-filters-10865353/
- https://www.theregister.com/security/2026/09/04/ascii-smuggling-isnt-just-an-ai-security-risk/5294595





