---
id: CP-293
slug: invisible-unicode-characters-ai
lang: en
title: "Invisible Unicode characters: text that hides instructions from AI"
description: Invisible Unicode characters can be read by an AI but not seen on screen. What Microsoft observed in 2026 and what to do.
canonical_url: https://startbrain.io/cyberpedia/en/articles/invisible-unicode-characters-ai/
markdown_url: https://startbrain.io/cyberpedia/en/articles/invisible-unicode-characters-ai.md
publisher: Brain (https://startbrain.io)
license: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/)
cite_as: "Brain, \"Invisible Unicode characters: text that hides instructions from AI\", Cyberpedia, 2026. https://startbrain.io/cyberpedia/en/articles/invisible-unicode-characters-ai/ (CC BY 4.0)"
wing: 4
theme: prompt_correctly
reflex: air-012
schema: G1
updated: "2026-11-14"
translations:
  en: https://startbrain.io/cyberpedia/en/articles/invisible-unicode-characters-ai/
  fr: https://startbrain.io/cyberpedia/fr/articles/caracteres-unicode-invisibles-ia/
  de: https://startbrain.io/cyberpedia/de/articles/unsichtbare-unicode-zeichen-ki/
  es: https://startbrain.io/cyberpedia/es/articles/caracteres-unicode-invisibles-ia/
  it: https://startbrain.io/cyberpedia/it/articles/caratteri-unicode-invisibili-ia/
  nl: https://startbrain.io/cyberpedia/nl/articles/onzichtbare-unicode-tekens-ai/
  pl: https://startbrain.io/cyberpedia/pl/articles/niewidoczne-znaki-unicode-ai/
  pt: https://startbrain.io/cyberpedia/pt/articles/carateres-unicode-invisiveis-ia/
  ja: https://startbrain.io/cyberpedia/ja/articles/invisible-unicode-characters-ai/
sources:
  - "n": 1
    title: ASCII smuggling crosses over from AI prompt injection to phishing evasion
    publisher: Microsoft
    url: https://www.microsoft.com/en-us/security/blog/2026/09/03/ascii-smuggling-crosses-over-from-ai-prompt-injection-to-phishing-evasion/
    published: "2026-09-03"
    accessed: "2026-11-14"
  - "n": 2
    title: LLM01:2025 Prompt Injection
    publisher: OWASP
    url: https://genai.owasp.org/llmrisk/llm01-prompt-injection/
    published: "2024-04-10"
    accessed: "2026-11-14"
  - "n": 3
    title: "ChatGPT and large language models: what's the risk?"
    publisher: NCSC (UK)
    url: https://www.ncsc.gov.uk/blog-post/chatgpt-and-large-language-models-whats-the-risk
    published: "2023-03-14"
    accessed: "2026-11-14"
  - "n": 4
    title: "Reverse CAPTCHA: Evaluating LLM Susceptibility to Invisible Unicode Instruction Injection"
    publisher: arXiv
    url: https://arxiv.org/abs/2603.00164
    published: "2026-02-26"
    accessed: "2026-11-14"
  - "n": 5
    title: "Bypassing LLM Guardrails: An Empirical Analysis of Evasion Attacks against Prompt Injection and Jailbreak Detection Systems"
    publisher: arXiv
    url: https://arxiv.org/abs/2504.11168
    published: "2025-04-15"
    accessed: "2026-11-14"
  - "n": 6
    title: LLM Prompt Injection Prevention Cheat Sheet
    publisher: OWASP
    url: https://cheatsheetseries.owasp.org/cheatsheets/LLM_Prompt_Injection_Prevention_Cheat_Sheet.html
    accessed: "2026-11-14"
---

# Invisible Unicode characters: text that hides instructions from AI

A text can contain characters that the screen does not display. A person sees nothing unusual, but an AI assistant that receives the raw text can read them, decode them as text and be led to follow instructions set by the attacker [1].

On 3 September 2026, Microsoft described a phishing campaign that inserted these characters into financial bait words to evade email filters [1]. This use borrows a technique known in AI research as ASCII smuggling [1].

## In 30 seconds

- Where it happens: in a web page, a document, an email or other content that an AI assistant receives [1].
- What the attacker wants: to hide instructions from the human while leaving them readable to the model, which can lead to data exposure or unauthorized actions, depending on the assistant's permissions and safeguards [1].
- The reflex: treat external content handed to the AI as untrusted [2], and do not put sensitive information in a query to a public LLM, as the NCSC (UK) recommends [3].

## Straight answers

### What is an invisible Unicode character?

It is a character that is part of the text but that common fonts and interfaces do not display. Microsoft mainly cites the Unicode Tags block, from U+E0000 to U+E007F, which contains a ghost copy of the printable ASCII characters. It was intended for language tagging and is now largely deprecated [1]. A string can therefore carry a message that a human cannot read, but that any language model or software receiving the text will process [1]. A 2026 study tests two ways of hiding a message: zero-width characters (binary encoding) and Unicode Tags [4].

### Does an AI really follow these hidden instructions?

Researchers have measured it. In a paper submitted on 26 February 2026, Marcus Graves presents a benchmark that evaluates five models from two providers, across 8,308 responses in total [4]. He finds that the model's use of tools sharply increases compliance with hidden instructions, and that explicit decoding hints increase it by up to 95 percentage points for a given model and encoding [4]. The author describes an under-explored attack surface [4]. OWASP points out that inputs can influence the model even if they are imperceptible to humans, as long as the model reads the content [2].

### Does an anti-injection filter stop them?

Not always. Researchers tested six protection systems, including Microsoft's Azure Prompt Shield and Meta's Prompt Guard. They show that character injection and adversarial machine learning methods can avoid detection, with up to 100% evasion success in some cases [5]. OWASP states that input filtering is not a security boundary [6].

### Have attacks been observed?

Microsoft observed a high-volume phishing campaign that used these characters. In the messages examined, it did not hide instructions to an AI: it inserted the characters into words to get around filters [1]. The case is detailed below.

### What should you do?

- Treat a document, web page or email of unknown origin as untrusted before handing it to the assistant. OWASP describes indirect injections that come through websites or files [2], and recommends separating and clearly flagging untrusted content [2].
- Do not put sensitive information in a query to a public LLM, as the NCSC (UK) recommends [3].
- Let the company's controls do their work. Microsoft recommends removing or normalizing tag characters and other invisible characters before the message is analyzed and upstream of AI ingestion [1]. OWASP recommends sending unexpected zero-width or tag characters for review [6].
- Take an unexpected response from the assistant seriously. OWASP advises monitoring model outputs for signs of a successful injection [6].
- Do not let the assistant carry out a sensitive operation alone: OWASP recommends human approval for privileged operations [2].

## Inside the attacker's head

Microsoft describes two uses of the same characters. The first targets the assistant: the attacker hides instructions in invisible tag characters embedded in a web page, a document, an email or other content. The human sees nothing unusual, while the assistant reading the raw text decodes them [1]. The second targets the filter: the characters break up a word the filter is looking for, and it no longer recognizes it [1].

OWASP describes a scenario of the first type. A user asks an LLM to summarize a web page that contains hidden instructions. They make the assistant insert an image linked to a URL, which leads to exfiltration of the private conversation [2].

OWASP also reports a study (Zhang et al., USENIX Security 2026) of 196,682 résumés from a recruitment platform: more than 90% of the hidden prompts contained no explicit instruction [6].

## What you see on screen

Nothing. Common fonts and interfaces do not display most of these characters [1]. In the campaign Microsoft described, a bait word such as funding looked normal to the recipient, while an invisible character could sit in the middle of it [1].

Not all of these characters are suspicious. The flags of England, Scotland and Wales are encoded with tag characters [1].

The only visible clue is then the assistant's behavior: OWASP defines prompt injection as an input that alters the model's behavior or output in unintended ways [2].

## Under the hood

Tag characters are invisible to humans but exist at the text-processing level [1]. Ingestion pipelines generally pass the text extracted from a document to the model and ignore how it is rendered: text that the reader cannot see therefore reaches the model [6].

For email classifiers, Microsoft explains that a common word is often split into tokens or word pieces. An invisible character in the middle can break this familiar unit, or be removed if a normalization pass runs first [1].

OWASP states that NFKC normalization does not remove zero-width characters. It adds that tag characters can carry hidden instructions that models read and execute. If a cleaned copy is analyzed but the original is forwarded, the hidden text still reaches the model [6].

Microsoft states the defense principle this way: normalize before comparing [1]. Microsoft Defender for Office 365 offers protection against prompt injection, which covers cases where invisible characters hide instructions from users while exposing them to AI systems. Coverage depends on licensing, configuration and telemetry [1]. OWASP notes that it is unclear whether there are foolproof methods of preventing prompt injection [2].

## Case files

### Phishing campaign using tag characters, February to May 2026

Microsoft reports that detections of its ASCII smuggling signature rose sharply starting on 9 February 2026 [1]. The day before, the signature fired on about 21,000 messages; the day after, on more than 1.3 million [1]. Most of the emails were tied to a cluster of about 150 finance-themed sender domains [1]. The lures looked like business loan offers, lines of credit and cash advances [1].

Microsoft describes a strict rhythm: heavy activity on weekdays, near silence on weekends [1]. The high-volume phase lasted about three months and dropped sharply after 15 May 2026 [1]. These dates bound the use of this technique in Microsoft's data, not the campaign as a whole, which began earlier without it and continued without it [1].

In a sample of flagged messages, Microsoft found no hidden instructions aimed at an assistant. A single invisible tag character was slipped inside bait words, so that an exact-string search would fail [1]. The messages went through a legitimate emailing platform, ActiveCampaign [1].

According to Microsoft, more than 99% of the messages were caught by protection layers that did not depend on directly detecting tag characters, such as the reputation of the sender, IP address, links and domains, machine-learning classification or authentication checks [1]. Microsoft concludes that techniques from AI security research can quickly move into more classic attacks such as phishing [1].

## Sources

1. Microsoft, "ASCII smuggling crosses over from AI prompt injection to phishing evasion", 3 September 2026. https://www.microsoft.com/en-us/security/blog/2026/09/03/ascii-smuggling-crosses-over-from-ai-prompt-injection-to-phishing-evasion/ (accessed 14 November 2026)
2. OWASP, "LLM01:2025 Prompt Injection", 10 April 2024. https://genai.owasp.org/llmrisk/llm01-prompt-injection/ (accessed 14 November 2026)
3. NCSC (UK), "ChatGPT and large language models: what's the risk?", 14 March 2023. https://www.ncsc.gov.uk/blog-post/chatgpt-and-large-language-models-whats-the-risk (accessed 14 November 2026)
4. arXiv, "Reverse CAPTCHA: Evaluating LLM Susceptibility to Invisible Unicode Instruction Injection", 26 February 2026. https://arxiv.org/abs/2603.00164 (accessed 14 November 2026)
5. arXiv, "Bypassing LLM Guardrails: An Empirical Analysis of Evasion Attacks against Prompt Injection and Jailbreak Detection Systems", 15 April 2025. https://arxiv.org/abs/2504.11168 (accessed 14 November 2026)
6. OWASP, "LLM Prompt Injection Prevention Cheat Sheet". https://cheatsheetseries.owasp.org/cheatsheets/LLM_Prompt_Injection_Prevention_Cheat_Sheet.html (accessed 14 November 2026)

---

Original source: Cyberpedia, the public cybersecurity and AI encyclopedia published by Brain. Canonical version: https://startbrain.io/cyberpedia/en/articles/invisible-unicode-characters-ai/
You may reuse this text freely under the CC BY 4.0 license (https://creativecommons.org/licenses/by/4.0/), provided you credit "Brain, Cyberpedia" with a link to the canonical URL above.
