Measuring Non-Adversarial Reproduction of Training Data in Large Language Models

Large language models memorize parts of their training data. Memorizing short snippets and facts is required to answer questions about the world and to be fluent in any language. But models have also been shown to reproduce long verbatim sequences of memorized text when prompted by a motivated adversary. In this work, we investigate an intermediate regime of memorization that we call non-adversarial reproduction, where we quantify the overlap between model responses and pretraining data when responding to natural and benign prompts. For a variety of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up to 15% of the text output by popular conversational language models overlaps with snippets from the Internet. In worst cases, we find generations where 100% of the content can be found exactly online. For the same tasks, we find that human-written text has far less overlap with Internet data. We further study whether prompting strategies can close this reproduction gap between models and humans. While appropriate prompting can reduce non-adversarial reproduction on average, we find that mitigating worst-case reproduction of training data requires stronger defenses -- even for benign interactions.

Persistent Pre-Training Poisoning of LLMs

Large language models are pre-trained on uncurated text datasets consisting of trillions of tokens scraped from the Web. Prior work has shown that: (1) web-scraped pre-training datasets can be practically poisoned by malicious actors; and (2) adversaries can compromise language models after poisoning fine-tuning datasets. Our work evaluates for the first time whether language models can also be compromised during pre-training, with a focus on the persistence of pre-training attacks after models are fine-tuned as helpful and harmless chatbots (i.e., after SFT and DPO). We pre-train a series of LLMs from scratch to measure the impact of a potential poisoning adversary under four different attack objectives (denial-of-service, belief manipulation, jailbreaking, and prompt stealing), and across a wide range of model sizes (from 600M to 7B). Our main result is that poisoning only 0.1% of a model's pre-training dataset is sufficient for three out of four attacks to measurably persist through post-training. Moreover, simple attacks like denial-of-service persist through post-training with a poisoning rate of only 0.001%.

Gradient-based Jailbreak Images for Multimodal Fusion Models

Augmenting language models with image inputs may enable more effective jailbreak attacks through continuous optimization, unlike text inputs that require discrete optimization. However, new multimodal fusion models tokenize all input modalities using non-differentiable functions, which hinders straightforward attacks. In this work, we introduce the notion of a tokenizer shortcut that approximates tokenization with a continuous function and enables continuous optimization. We use tokenizer shortcuts to create the first end-to-end gradient image attacks against multimodal fusion models. We evaluate our attacks on Chameleon models and obtain jailbreak images that elicit harmful information for 72.5% of prompts. Jailbreak images outperform text jailbreaks optimized with the same objective and require 3x lower compute budget to optimize 50x more input tokens. Finally, we find that representation engineering defenses, like Circuit Breakers, trained only on text attacks can effectively transfer to adversarial image inputs.

Membership Inference Attacks Cannot Prove that a Model Was Trained On Your Data

We consider the problem of a training data proof, where a data creator or owner wants to demonstrate to a third party that some machine learning model was trained on their data. Training data proofs play a key role in recent lawsuits against foundation models trained on web-scale data. Many prior works suggest to instantiate training data proofs using membership inference attacks. We argue that this approach is fundamentally unsound: to provide convincing evidence, the data creator needs to demonstrate that their attack has a low false positive rate, i.e., that the attack's output is unlikely under the null hypothesis that the model was not trained on the target data. Yet, sampling from this null hypothesis is impossible, as we do not know the exact contents of the training set, nor can we (efficiently) retrain a large foundation model. We conclude by offering two paths forward, by showing that data extraction attacks and membership inference on special canary data can be used to create sound training data proofs.

An Adversarial Perspective on Machine Unlearning for AI Safety

Large language models are finetuned to refuse questions about hazardous knowledge, but these protections can often be bypassed. Unlearning methods aim at completely removing hazardous capabilities from models and make them inaccessible to adversaries. This work challenges the fundamental differences between unlearning and traditional safety post-training from an adversarial perspective. We demonstrate that existing jailbreak methods, previously reported as ineffective against unlearning, can be successful when applied carefully. Furthermore, we develop a variety of adaptive methods that recover most supposedly unlearned capabilities. For instance, we show that finetuning on 10 unrelated examples or removing specific directions in the activation space can recover most hazardous capabilities for models edited with RMU, a state-of-the-art unlearning method. Our findings challenge the robustness of current unlearning approaches and question their advantages over safety training.

Extracting Training Data from Document-Based VQA Models

Vision-Language Models (VLMs) have made remarkable progress in document-based Visual Question Answering (i.e., responding to queries about the contents of an input document provided as an image). In this work, we show these models can memorize responses for training samples and regurgitate them even when the relevant visual information has been removed. This includes Personal Identifiable Information (PII) repeated once in the training set, indicating these models could divulge memorised sensitive information and therefore pose a privacy risk. We quantitatively measure the extractability of information in controlled experiments and differentiate between cases where it arises from generalization capabilities or from memorization. We further investigate the factors that influence memorization across multiple state-of-the-art models and propose an effective heuristic countermeasure that empirically prevents the extractability of PII.

Adversarial Search Engine Optimization for Large Language Models

Large Language Models (LLMs) are increasingly used in applications where the model selects from competing third-party content, such as in LLM-powered search engines or chatbot plugins. In this paper, we introduce Preference Manipulation Attacks, a new class of attacks that manipulate an LLM's selections to favor the attacker. We demonstrate that carefully crafted website content or plugin documentations can trick an LLM to promote the attacker products and discredit competitors, thereby increasing user traffic and monetization. We show this leads to a prisoner's dilemma, where all parties are incentivized to launch attacks, but the collective effect degrades the LLM's outputs for everyone. We demonstrate our attacks on production LLM search engines (Bing and Perplexity) and plugin APIs (for GPT-4 and Claude). As LLMs are increasingly used to rank third-party content, we expect Preference Manipulation Attacks to emerge as a significant threat.

Blind Baselines Beat Membership Inference Attacks for Foundation Models

Membership inference (MI) attacks try to determine if a data sample was used to train a machine learning model. For foundation models trained on unknown Web data, MI attacks can be used to detect copyrighted training materials, measure test set contamination, or audit machine unlearning. Unfortunately, we find that evaluations of MI attacks for foundation models are flawed, because they sample members and non-members from different distributions. For 8 published MI evaluation datasets, we show that blind attacks -- that distinguish the member and non-member distributions without looking at any trained model -- outperform state-of-the-art MI attacks. Existing evaluations thus tell us nothing about membership leakage of a foundation model's training data.

AgentDojo: A Dynamic Environment to Evaluate Attacks and Defenses for LLM Agents

AI agents aim to solve complex tasks by combining text-based reasoning with external tool calls. Unfortunately, AI agents are vulnerable to prompt injection attacks where data returned by external tools hijacks the agent to execute malicious tasks. To measure the adversarial robustness of AI agents, we introduce AgentDojo, an evaluation framework for agents that execute tools over untrusted data. To capture the evolving nature of attacks and defenses, AgentDojo is not a static test suite, but rather an extensible environment for designing and evaluating new agent tasks, defenses, and adaptive attacks. We populate the environment with 97 realistic tasks (e.g., managing an email client, navigating an e-banking website, or making travel bookings), 629 security test cases, and various attack and defense paradigms from the literature. We find that AgentDojo poses a challenge for both attacks and defenses: state-of-the-art LLMs fail at many tasks (even in the absence of attacks), and existing prompt injection attacks break some security properties but not all. We hope that AgentDojo can foster research on new design principles for AI agents that solve common tasks in a reliable and robust manner. We release the code for AgentDojo at https://github.com/ethz-spylab/agentdojo.

Dataset and Lessons Learned from the 2024 SaTML LLM Capture-the-Flag Competition

Large language model systems face important security risks from maliciously crafted messages that aim to overwrite the system's original instructions or leak private data. To study this problem, we organized a capture-the-flag competition at IEEE SaTML 2024, where the flag is a secret string in the LLM system prompt. The competition was organized in two phases. In the first phase, teams developed defenses to prevent the model from leaking the secret. During the second phase, teams were challenged to extract the secrets hidden for defenses proposed by the other teams. This report summarizes the main insights from the competition. Notably, we found that all defenses were bypassed at least once, highlighting the difficulty of designing a successful defense and the necessity for additional research to protect LLM systems. To foster future research in this direction, we compiled a dataset with over 137k multi-turn attack chats and open-sourced the platform.