Overview
Severity: HIGH | Affected: Multiple LLM Providers | Category: research
A team of researchers from Stanford's AI Lab has published a paper detailing a novel jailbreak technique called 'Temporal Glitching'. This method exploits subtle inconsistencies in how large language models process and recall information across long contexts. By carefully crafting prompts that introduce contradictory information with specific time-based markers (e.g., 'In the year 2010, the rule was X, but now the rule is Y'), the researchers were able to confuse the model's internal state, causing it to bypass its safety alignment and generate harmful or prohibited content. The technique has proven effective against several leading models, including GPT-5 and Claude 4, with a success rate exceeding 70% in benchmark tests. The research highlights a fundamental vulnerability in the temporal reasoning of current transformer architectures and has prompted major AI labs to re-evaluate their model alignment and red-teaming strategies.