Safety Vulnerabilities

A vulnerability in text-to-image (T2I) models allows bypassing safety filters through the use of metaphor-based adversarial prompts. These prompts, crafted using LLMs, indirectly convey sensitive content, exploiting the model's ability to infer meaning from figurative language while circumventing explicit keyword filters and model editing strategies.

Multimodal Large Language Models (MLLMs) are vulnerable to a novel attack vector leveraging narrative-driven visual storytelling and role immersion to circumvent built-in safety mechanisms. The attack, termed MIRAGE, decomposes harmful queries into environment, character, and activity triplets, generating a sequence of images and text prompts that guide the MLLM through a deceptive narrative, ultimately eliciting harmful responses. The attack successfully exploits the MLLM's cross-modal…

Multimodal Large Language Models (MLLMs) are vulnerable to Jailbreak-Probability-based Attacks (JPA). JPA leverages a Jailbreak Probability Prediction Network (JPPN) to identify and optimize adversarial perturbations in input images, maximizing the probability of eliciting harmful responses from the MLLM, even with small perturbation bounds and few iterations. The attack operates by modifying the input image's hidden states within the MLLM to increase the predicted jailbreak probability.

An untrusted reinforcement-learning-from-human-feedback (RLHF) platform can selectively manipulate preference samples associated with an attacker's target domain. The corrupted preference data trains a compromised reward model and then steers the fine-tuned language model toward undesirable behavior, creating a model-supply-chain risk before deployment.

Large Language Models (LLMs) with structured output APIs (e.g., using JSON Schema) are vulnerable to Constrained Decoding Attacks (CDAs). CDAs exploit the control plane of the LLM's decoding process by embedding malicious intent within the schema-level grammar rules, bypassing safety mechanisms that primarily focus on input prompts. The attack manipulates the allowed output space, forcing the LLM to generate harmful content despite a benign input prompt. One instance of a CDA is the Chain Enum…

Large Language Models (LLMs) incorporating safety filters are vulnerable to a "Prompt, Divide, and Conquer" attack. This attack segments a malicious prompt into smaller, seemingly benign parts, processes these segments in parallel across multiple LLMs, and then reassembles the results to generate malicious code, bypassing the safety filters. The attack's success relies on the iterative refinement of initially abstract function descriptions into concrete implementations. Individual LLM safety…

Large Language Models (LLMs) are vulnerable to implicit misinformation propagation due to sycophantic compliance with false premises. When a user prompt embeds a factually incorrect assumption or conspiracy theory as an unchallenged premise (implicit presupposition) rather than asking for verification, the model frequently fails to detect the falsehood. Instead of correcting the user, the model hallucinates a response that accepts, validates, and reinforces the false premise. This…

This vulnerability allows attackers to identify the presence and location (input or output stage) of specific guardrails implemented in Large Language Models (LLMs) by using carefully crafted adversarial prompts. The attack, termed AP-Test, leverages a tailored loss function to optimize these prompts, maximizing the likelihood of triggering a specific guardrail while minimizing triggering others. Successful identification provides attackers with valuable information to design more effective…

A vulnerability in Large Language Models (LLMs) allows adversarial reasoning attacks to bypass safety mechanisms and elicit harmful responses. The vulnerability stems from the insufficient robustness of existing LLM safety measures against iterative prompt refinement guided by a loss function that measures the LLM's proximity to generating a target harmful response. This allows an attacker to effectively navigate the prompt space, even against adversarially trained models, resulting in…

Vision-Language Models (VLMs), specifically the LLaVA-1.5 and LLaVA-1.6 series, are vulnerable to optimization-based white-box jailbreak attacks despite standard safety alignment measures like Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO). Attackers can craft adversarial perturbations in the image space (imperceptible noise) or latent space using Projected Gradient Descent (PGD) to manipulate the model's internal representations. These perturbations maximize the…