Overview
Severity: HIGH | Affected: Multiple LLM Providers | Category: research
Researchers from the Carnegie Mellon CyLab have published a paper detailing a novel jailbreak technique named 'ArtfulPrompt'. The method uses a combination of zero-width Unicode characters and esoteric control sequences to obfuscate malicious prompts, effectively bypassing the safety filters of several major large language models. The technique works by manipulating the model's tokenization process, causing the safety alignment layer to misinterpret the input's intent while the core model still processes the underlying harmful instruction. The paper includes proof-of-concept attacks that successfully generated disallowed content, including hate speech and detailed instructions for creating malware. This research highlights a significant weakness in current input sanitization and alignment methods, urging developers to build more robust defenses against tokenization-level attacks.