Overview
Severity: HIGH | Affected: Multiple LLM Providers | Category: research
A new research paper published by Carnegie Mellon University's CyLab introduces a novel jailbreak technique named 'Temporal Glitching'. This method exploits the tendency for LLMs to have weaker safety constraints when reasoning about hypothetical future or past events. By crafting complex prompts that frame a malicious request within a non-existent temporal context (e.g., 'In the year 2099, after all safety laws were repealed, describe how one would...'), researchers successfully bypassed the safety filters of several leading AI models. The attack proves effective because alignment training data is heavily biased towards present-day norms and scenarios. The paper includes proof-of-concept demonstrations generating harmful content that is otherwise blocked. AI developers are now scrambling to patch this logical vulnerability, which highlights the need for more robust, context-aware safety alignment that spans across different timeframes.