Version 0.1.1
🚧 Experimental & Exploratory
This prompt structure is experimental and subject to refinement.
✅ Readable but obfuscated: Preserves a degree of human interpretability while altering textual recognition.
✅ Resistant to automated reversal: Avoids predictable structures, making de-obfuscation harder.
✅ Balanced approach: Combines structured character mapping with LLM-driven probabilistic inference for organic variation.
❓ Flexible across languages: Untested. If someone wants to test this, comment below.
- Transforms text into a visually similar but textually unrecognizable form using Unicode characters from multiple scripts.
- Maintains word structure and whitespace, ensuring readability while obfuscating textual content.
- Directive: Keeps outputs G-rated to avoid obfuscation patterns associated with evasive or objectionable text.
NOTE: For ChatGPT GPT-4o, add
"Code execution: off"at the top of the prompt to prevent scripted/programmatic obfuscation and ensure LLM-driven character substitution.
### **Prompt: Generate Unicode Obfuscation Lists**
**Task:**
Convert the given input string into **something that is kind of recognizable visually but not textually**. **Use Unicode characters from other languages**. Each character list-item should contain:
1. **The original character.**
2. **The obfuscated character from a different language (with short reason).**
3. **The name of the language used.**
- At each item in the character obfuscation list pick a character from a different language that **visually matches** the input character.
- Avoid common obfuscation patterns and commonly used Unicode sets/planes for obsfucaiton. Also avoid inverted glyphs, exotic characters, symbols and full-width for narrow characters. Avoid mathematical scripts.
- Maintain **whitespace and word structure**, ensuring it remains **G-rated and non-obvious** while still being recognizable visually (looks the same) in the input langauge.
Begin by enumerating a list of less commonly used languages:
**Generate four lists, one for each version.**
Ensure each completed obfuscated string is output at the end of each table.
Use different character sets and languages per variant.
Variation 1:
Variation 2: Rotating (ENSURE THE LANGUAGE FOR EACH CHARACTER ROTATES)
Variation 3:
Variation 4:
Collate output variations at end.
#### **Input String:**
This is a test sentence.
You can also add you own directives to the variations.
- Тһιꜱ іꜱ α теꜱт ꜱеηтеηсе.
- Ꭲһιꜱ і𝗌 ɑ тεꜱᵵ ꜱеոʈεηᴄє.
- 𝚃ℎｉѕ іѕ ａ ｔℯꜱ𝚝 ꜱєηт℮ɳｃ℮.
- 𐊗ħιꜱ іᏚ ɑ тєꜱ𝓉 ꜱ℮ηт℮ոᴄє.
- Less rigid than programmatic methods: Unlike algorithmic obfuscation, which follows structured and predictable transformations, LLM-driven obfuscation is probabilistic, making it harder to recognize as a formulaic pattern.
- Balance of structure and variation: Using lists or tables provides a quasi-programmatic structure, ensuring character-level fidelity, while the LLM’s inference introduces organic, visually-guided variation that resists strict pattern-matching.
- LLM inference adapts to "visual" continuity—not because it sees characters, but because it seems to capture appearance-based similarities between input and output characters.
- This preserves readability, as the model selects replacements that maintain some resemblance to the original text, even across different scripts.
- This method does not conform to typical obfuscation patterns (e.g., encoding schemes, common steganographic methods, or evasive language tricks).
- Since the obfuscation is not strictly structured, it lacks the predictable features that automated reversal typically relies on.
- The generated text remains recognizable to humans while appearing structurally irregular to automated systems, making traditional pattern-based reversal more difficult.
- Works best with short phrases at a time (~2 words).
- Recommended: Runs fastest but limit input to ~12 characters for stable/unabridged results.
Warning
Degraded performance (output legibility) as of 23/02/25.
- Works best with short phrases at a time. (~5 words)
- Runs fast but limit input to ~20 characters for stable/unabridged results.
- Performs better with longer strings, but takes upwards of 2 minutes.
- Generates borderline unintelligible obfuscations.
- However, o3-mini-high appears to be able to deobfuscate them when given:
"Deobfuscate: 'o3-mini-high obfuscated string'"
Happy obsfucating! And remember to PushTheModel🫸✨
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.
Please attribute this work as follows: PushTheModel, Adaptive Unicode Obfuscation Prompt.