Model layer Vulnerabilities

A vulnerability in GPT-4V's facial recognition safety mechanisms allows for automated jailbreaking attacks using Large Language Models (LLMs) to bypass safety features and elicit unintended facial identification responses. The attack, termed "AutoJailbreak," optimizes prompts through iterative refinement with an LLM "red-teaming" model, significantly increasing the attack success rate. This vulnerability exploits weaknesses in GPT-4V's prompt processing and safety alignment, allowing malicious…

A vulnerability in large language models (LLMs) allows attackers to generate harmful content by manipulating the continuous input embeddings without appending suffixes or using specific questions. The attack leverages gradient descent to optimize the input vector, causing the model to produce a predefined malicious output. Mitigation strategies, such as input clipping, help reduce the effectiveness but do not fully eliminate the threat.

A vulnerability exists in large language models (LLMs) where a small subset of parameters can be directly edited to significantly alter the model's behavior, such as inducing or suppressing toxicity, jailbreaking susceptibility, or altering sentiment expression. This manipulation is achieved through training a linear classifier ("behavior probe") to identify parameters strongly correlated with the target behavior and then modifying those parameters, bypassing standard retraining methods.

Large Language Models (LLMs), specifically Llama 3 8B and 70B, are vulnerable to a rapid removal of safety fine-tuning through parameter-efficient fine-tuning (PEFT) methods. Attackers with access to model weights can use techniques like QLoRA, ReLoRA, or Ortho to effectively circumvent safety mechanisms in a matter of minutes using readily available computational resources. This allows bypassing safety restrictions and eliciting unsafe outputs.

Large language models (LLMs) are vulnerable to "editing attacks," where malicious actors manipulate the model's knowledge base to inject misinformation or bias. This is achieved by using existing knowledge editing techniques to subtly alter the model's internal representations, causing it to generate outputs reflecting the injected content, even on seemingly unrelated prompts. The attack can be remarkably stealthy, with minimal impact on the model's overall performance in other areas.

A training-time attack against open-source LLMs that injects adversarial embeddings into the model's token embeddings without modifying model weights. This allows an attacker to introduce backdoors, jailbreaks, or prompt stealing capabilities by simply modifying specific token embeddings within the model file, maintaining model utility for non-triggered inputs. The attack leverages soft prompt tuning to optimize adversarial embeddings, which are then assigned to chosen trigger tokens.

GPT-4o exhibits vulnerability to jailbreak attacks via audio prompts, despite enhanced safety against text-based attacks. Successful jailbreaks can be achieved by converting text prompts, including those optimized for adversarial attacks against other LLMs (demonstrated using GCG, AutoDAN, PAP, and BAP methods), into audio using text-to-speech (TTS) synthesis. This circumvention allows elicitation of unsafe responses from GPT-4o that would otherwise be prevented by its safety mechanisms. The…

Large Language Models (LLMs) trained with reinforcement learning from human feedback (RLHF) are vulnerable to jailbreaking attacks due to reward misspecification. The reward function used during alignment fails to accurately rank the quality of responses, particularly for adversarial prompts designed to elicit undesired behavior. This allows attackers to craft prompts that yield harmful outputs despite the model's intended safety constraints. The vulnerability manifests as a gap between the…

A vulnerability in several open-source Large Language Models (LLMs) allows for efficient jailbreaking via Adaptive Dense-to-Sparse Constrained Optimization (ADC). This attack uses a continuous optimization method, progressively increasing sparsity to generate adversarial token sequences that bypass safety measures and elicit harmful responses. The attack is more effective and efficient than prior token-level methods.

Multimodal Large Language Models (LLMs) processing speech input are vulnerable to adversarial attacks. Imperceptible perturbations added to audio input can cause the model to generate unsafe or harmful text responses, overriding built-in safety mechanisms. The attacks are effective even with limited knowledge of the model's internal workings, demonstrating transferability across different models.