LLM Prompt Obfuscation
LLM Prompt Obfuscation is like writing a secret message in invisible ink or a code that only the computer can read. It is a trick used by attackers to hide malicious instructions inside a prompt so that people or security systems don't realize the computer is being told to do something harmful.
LLM Prompt Obfuscation involves techniques used to conceal malicious payloads within prompts or retrieval-augmented generation (RAG) data to evade detection. This includes visual tricks like hidden text or color-matching, embedding instructions within non-text data like image pixels or file metadata, and using encoding schemes like Base64 or ROT13 to bypass security filters and guardrails.
LLM Prompt Obfuscation is a class of adversarial evasion techniques designed to circumvent input sanitization, guardrails, and human oversight by decoupling the semantic intent of a prompt from its literal representation. This encompasses steganographic embedding of instructions within multi-modal artifacts (e.g., LSB manipulation in images, metadata injection), the exploitation of rendering-layer vulnerabilities (e.g., CSS-based text masking), and the application of obfuscation primitives (e.g., character-level encoding, token-level transformations) to neutralize signature-based or heuristic detection mechanisms while maintaining the model's ability to interpret the underlying malicious directive.