Model layer Vulnerabilities

Large Language Models (LLMs) used as safety judges are vulnerable to an "Emoji Attack," a prompt injection technique that leverages token segmentation bias. Inserting emojis within tokens alters sub-token embeddings, misleading the judge LLM into classifying harmful content as safe. The attack's effectiveness is amplified by strategically placing emojis to maximize the embedding discrepancy between sub-tokens and the original token.

A vulnerability exists in Federated Parameter-Efficient Fine-Tuning (FedPEFT) systems for large language models (LLMs). Malicious clients can exploit the PEFT mechanism (e.g., LoRA, (IA)³, LayerNorm) to inject adversarial training data, compromising the model's safety alignment even with a small percentage of trainable parameters and a minority of malicious participants. The attack, termed "PEFT-as-an-Attack" (PaaA), circumvents the LLM's safety guardrails, causing it to generate harmful…

Large Language Models (LLMs) are vulnerable to a jailbreak attack that leverages the model's ability to generate diverse and obfuscated prompts to bypass safety constraints. The attack exploits the model's capacity to deviate from prior context, rendering existing safety training ineffective. The attacker uses a multi-stage process involving diversification (generating prompts significantly different from previous attempts) and obfuscation (obscuring sensitive words/phrases) to elicit harmful…

Large language models (LLMs) are vulnerable to jailbreak attacks exploiting nonlinear features within prompt encodings. These features, not detectable by linear methods, allow adversaries to reliably elicit harmful outputs despite safety training. Different attack methods leverage distinct nonlinear features, limiting the transferability of detection and mitigation techniques.

The LLaMA-2-7b-chat large language model (LLM) is vulnerable to a prompt-driven attack, termed DROJ (Directed Representation Optimization Jailbreak), that optimizes prompts at the embedding level to circumvent safety mechanisms and elicit harmful responses. The attack shifts the hidden representations of harmful queries away from the model's refusal direction, leading to a high attack success rate even with safety prompts in place. While the model may not refuse, responses may be repetitive…

LLMStinger uses a reinforcement-learning loop to fine-tune an attacker model that generates adversarial suffixes for jailbreak prompts. Because the approach does not require white-box access to the target, it can adapt existing attacks for both open- and closed-source assistants and bypass otherwise effective refusal behavior.

Large Language Models (LLMs) are vulnerable to a novel prompt injection attack using universal and context-independent triggers. These triggers, once discovered for a specific LLM, allow precise control over the model's output regardless of the prompt context or desired output content, enabling adversaries to force the generation of arbitrary text. The attack utilizes a gradient-based optimization technique to discover these triggers.

A vulnerability in multi-modal large language models (MLLMs) allows attackers to bypass safety mechanisms and elicit harmful responses using a memory-efficient zeroth-order optimization technique. The attack, termed Zer0-Jack, leverages simultaneous perturbation stochastic approximation (SPSA) with patch coordinate descent to generate malicious image inputs, even without access to the model's internal parameters (black-box setting).

Large Language Models (LLMs) are vulnerable to attention-based jailbreak attacks. Attackers can craft prompts that strategically divert the LLM's attention away from sensitive words, causing the model to overlook malicious intent and generate harmful content. This occurs by leveraging the LLM's attention mechanism to focus on benign parts of the prompt while embedding harmful queries within a seemingly harmless context. The success of the attack is correlated with specific attention…

Large Language Models (LLMs) are vulnerable to jailbreaking attacks that manipulate attention scores to redirect the model's focus away from safety protocols. The AttnGCG attack method increases the attention score on adversarial suffixes within the input prompt, causing the model to prioritize the malicious content over safety guidelines, leading to the generation of harmful outputs.