Tensor Trust: Interpretable Prompt Injection Attacks from an Online Game

While Large Language Models (LLMs) are increasingly being used in real-world applications, they remain vulnerable to prompt injection attacks: malicious third party prompts that subvert the intent of the system designer. To help researchers study this problem, we present a dataset of over 126,000 prompt injection attacks and 46,000 prompt-based "defenses" against prompt injection, all created by players of an online game called Tensor Trust. To the best of our knowledge, this is currently the largest dataset of human-generated adversarial examples for instruction-following LLMs. The attacks in our dataset have a lot of easily interpretable stucture, and shed light on the weaknesses of LLMs. We also use the dataset to create a benchmark for resistance to two types of prompt injection, which we refer to as prompt extraction and prompt hijacking. Our benchmark results show that many models are vulnerable to the attack strategies in the Tensor Trust dataset. Furthermore, we show that some attack strategies from the dataset generalize to deployed LLM-based applications, even though they have a very different set of constraints to the game. We release all data and source code at https://tensortrust.ai/paper

Image Hijacks: Adversarial Images can Control Generative Models at Runtime

Are foundation models secure from malicious actors? In this work, we focus on the image input to a vision-language model (VLM). We discover image hijacks, adversarial images that control generative models at runtime. We introduce Behaviour Matching, a general method for creating image hijacks, and we use it to explore three types of attacks. Specific string attacks generate arbitrary output of the adversary's choice. Leak context attacks leak information from the context window into the output. Jailbreak attacks circumvent a model's safety training. We study these attacks against LLaVA, a state-of-the-art VLM based on CLIP and LLaMA-2, and find that all our attack types have above a 90% success rate. Moreover, our attacks are automated and require only small image perturbations. These findings raise serious concerns about the security of foundation models. If image hijacks are as difficult to defend against as adversarial examples in CIFAR-10, then it might be many years before a solution is found -- if it even exists.

Image Hijacking: Adversarial Images can Control Generative Models at Runtime

Are foundation models secure from malicious actors? In this work, we focus on the image input to a vision-language model (VLM). We discover image hijacks, adversarial images that control generative models at runtime. We introduce Behavior Matching, a general method for creating image hijacks, and we use it to explore three types of attacks. Specific string attacks generate arbitrary output of the adversary's choosing. Leak context attacks leak information from the context window into the output. Jailbreak attacks circumvent a model's safety training. We study these attacks against LLaVA-2, a state-of-the-art VLM based on CLIP and LLaMA-2, and find that all our attack types have above a 90\% success rate. Moreover, our attacks are automated and require only small image perturbations. These findings raise serious concerns about the security of foundation models. If image hijacks are as difficult to defend against as adversarial examples in CIFAR-10, then it might be many years before a solution is found -- if it even exists.

Bigger&Faster: Two-stage Neural Architecture Search for Quantized Transformer Models

Neural architecture search (NAS) for transformers has been used to create state-of-the-art models that target certain latency constraints. In this work we present Bigger&Faster, a novel quantization-aware parameter sharing NAS that finds architectures for 8-bit integer (int8) quantized transformers. Our results show that our method is able to produce BERT models that outperform the current state-of-the-art technique, AutoTinyBERT, at all latency targets we tested, achieving up to a 2.68% accuracy gain. Additionally, although the models found by our technique have a larger number of parameters than their float32 counterparts, due to their parameters being int8, they have significantly smaller memory footprints.