A proof-of-concept tool demonstrating invisible adversarial injection attacks against LLMs using Unicode steganography. Hides malicious instructions in zero-width characters and template engine metadata to bypass safety filters.
python proof-of-concept jailbreak cybersecurity steganography red-team security-research ai-security llm prompt-injection llm-security adversarial-ai unicode-steganography
-
Updated
Oct 13, 2026 - Python