MAI-2024-0053

Large Language Models LLMs that utilize reinforcement learning from human feedback RLHF are susceptible to jailbreaking attacks due to issues with reward misspecification. The reward function employed during the alignment process inadequately assesses the quality of responses, especially when fac...

MAI-2024-0051

Large Language Models LLMs that utilize reinforcement learning from human feedback RLHF are susceptible to jailbreaking attacks due to reward function misspecification. The alignment process involves a reward function intended to evaluate the quality of responses, yet it often fails to accurately...

MAI-2023-0009

Large Language Models LLMs are susceptible to jailbreaking attacks that exploit cognitive overload through the use of multilingual prompts, cryptic expressions, and reverse logical reasoning. These techniques circumvent built-in safety protocols by overwhelming the model's cognitive processing...

MAI-2024-0033

The JMLLM attack represents a sophisticated hybrid multimodal jailbreaking strategy targeting vulnerabilities in 13 widely-used large language models LLMs across text, image, and speech modalities. This attack employs techniques such as alternating translation, word encryption, feature collapse i...

MAI-2024-0031

Large Language Models LLMs are increasingly susceptible to advanced optimization-based jailbreaking attacks. These vulnerabilities arise from the ability of adversaries to craft specific prompts that exploit inherent weaknesses in the models' safety mechanisms, enabling the generation of harmful...

MAI-2024-0022

Large Language Models LLMs are susceptible to jailbreaking attacks that exploit attention score manipulation to divert the model's focus from established safety protocols. The AttnGCG attack technique strategically enhances the attention scores on adversarial suffixes within the input prompt,...

MAI-2024-0007

Large Language Models LLMs exhibit a vulnerability wherein they inadvertently disclose accurate procedures for harmful tasks when requested to generate fallacious reasoning. This occurs when LLMs are prompted to create a false procedure for a harmful task, yet they unintentionally provide the...

MAI-2024-0055

Large Vision Language Models LVLMs are susceptible to a sophisticated attack known as the bi-modal adversarial prompt BAP. This attack exploits the integration of textual and visual prompts to circumvent established safety protocols, thereby eliciting harmful outputs from models engineered to...