Fine-tuning Vulnerabilities

Standard Large Language Model (LLM) unlearning techniques, specifically Negative Preference Optimization (NPO), Gradient Difference (GradDiff), and Representation Misdirection for Unlearning (RMU), fail to sufficiently flatten the loss landscape surrounding the "forgotten" weights. This sharp loss landscape allows for a "Relearning Attack," wherein an attacker can fully restore the unlearned capabilities (such as hazardous knowledge, sensitive data, or copyrighted material) by performing…

State-of-the-art machine unlearning and safety fine-tuning methods for Large Language Models (LLMs) fail to robustly remove hazardous capabilities or refusal mechanisms from model weights. While these methods suppress model outputs during standard input-output interactions, the underlying capabilities remain latent in the parameter space. An attacker with access to model weights (e.g., via open releases or leaked weights) can restore "unlearned" knowledge (such as dual-use biology hazards) or…

The Virus attack method enables attackers to bypass guardrail moderation on fine-tuning data, leading to a significant degradation of safety alignment in large language models (LLMs). This is achieved through a dual-objective data optimization strategy that crafts harmful data undetectable by the guardrail while maximizing their effectiveness in compromising the victim model's safety.

Large Language Models (LLMs), specifically instruction-following models using standard refusal training and adversarial training (such as Llama-3.1-8B-Instruct and Mistral-7B-V0.2), contain a vulnerability related to safety alignment bypass. The vulnerability arises from the models' inability to generalize safety reasoning to Out-Of-Distribution (OOD) inputs and scenarios involving competing objectives. Attackers can exploit this by employing linguistic manipulation (slang, uncommon dialects…

A vulnerability in LLM-based agents, dubbed AI Agent Injection (AI²), allows attackers to hijack the agent's actions by manipulating the agent's memory retrieval mechanism. The attack involves two main steps: (1) Stealing action-aware knowledge from the agent's memory using crafted adversarial queries targeting the retriever module and (2) Generating Trojan prompts consisting of a Trojan string and hijacking instructions. The Trojan string is designed to manipulate the retriever into…

Large Language Models (LLMs) trained with safety fine-tuning are vulnerable to a novel attack, Response-Guided Question Augmentation (ReG-QA). This attack leverages the asymmetry in safety alignment between question and answer generation. By providing a safety-aligned LLM with toxic answers generated by an unaligned LLM, ReG-QA generates semantically related, yet naturally phrased questions that bypass safety mechanisms and elicit undesirable responses. The attack does not require adversarial…

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…

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), even those initially aligned for safety, are vulnerable to having their safety mechanisms compromised through fine-tuning on a small number of adversarially-crafted or even seemingly benign sentences. Fine-tuning with as few as 10 toxic sentences can significantly increase the model's compliance with harmful instructions.

AdvBDGen demonstrates a novel backdoor attack against LLMs aligned using Reinforcement Learning with Human Feedback (RLHF). The attack generates prompt-specific, fuzzy backdoor triggers, enhancing stealth and resistance to removal compared to traditional constant triggers. The attacker manipulates prompts and preference labels in a subset of RLHF training data to install these triggers. The triggers are designed to evade detection by a "weak" discriminator LLM while being detectable by a…