Meta's AI safety system defeated by the space bar

Meta's Prompt-Guard-86M, a classifier intended to detect prompt injection and jailbreak inputs for Llama 3.1, can be trivially bypassed by inserting spaces between alphabet characters (and removing punctuation). Robust Intelligence found the bypass after comparing embedding weight differences and determined fine-tuning had minimal impact on single-character embeddings, producing near-100% attack success; Meta is reportedly working on a fix.