Overview
Severity: HIGH | Affected: Multiple LLM Providers | Category: research
A paper from Carnegie Mellon University's CyLab has introduced a novel jailbreak technique named 'Semantic Jitter'. This method circumvents the safety filters of major large language models by introducing subtle, algorithmically-generated noise into user prompts. The noise characters, often visually imperceptible or appearing as typos, are strategically placed to disrupt the model's safety alignment layer without altering the core semantic meaning of the malicious instruction. The research demonstrated a success rate of over 85% in generating harmful content from several leading commercial and open-source models, a significant increase over previous techniques. The researchers followed a responsible disclosure protocol, giving top AI labs a 90-day window to patch the vulnerability before publication. This discovery exposes the brittleness of current alignment methods and underscores the need for more robust, context-aware defense mechanisms against adversarial inputs.