Prompt obfuscation is the practice of disguising harmful or policy-violating instructions so they look like ordinary text and are harder for automated filters to detect. Instead of obvious tricks such as Base64, unusual symbols, or hidden characters, the attacker uses fluent prose, indirect phrasing, or layered wording to conceal the real intent of the prompt.
This matters because many AI systems rely on lightweight guardrails to decide whether input is safe before a more capable model or tool runs. If the first screen is too shallow, obfuscated prompts can slip through and trigger prompt injection, unsafe generation, or unintended actions by an agent. Defenders reduce this risk by using layered checks, separating instructions from data, monitoring tool use and outputs, and limiting system privileges so a single missed prompt does not lead to full compromise.