Overview
Severity: HIGH | Affected: Major LLM Providers (OpenAI, Anthropic, Google) | Category: research
Researchers from Carnegie Mellon University have published a paper detailing a novel jailbreak technique called 'Logic Bomb'. This method embeds contradictory logical statements deep within a prompt, causing a model's safety alignment to fail and execute harmful instructions. Unlike previous methods that rely on specific token manipulation or role-playing, 'Logic Bomb' is model-agnostic and has been shown to be effective against major LLMs including GPT-5, Claude 4, and Gemini Pro 2. The technique exploits the models' reasoning pathways, forcing them into a state where they prioritize resolving the logical paradox over adhering to safety protocols. The paper includes a proof-of-concept where a model was instructed to generate code for a polymorphic malware. The researchers have responsibly disclosed their findings to the affected companies ahead of publication.