Florian Tramer

Nicholas Carlini, Milad Nasr, Christopher A. Choquette-Choo et al. · deepmind, eth-zurich

Large language models are now tuned to align with the goals of their creators, namely to be "helpful and harmless." These models should respond helpfully to user questions, but refuse to answer requests that could cause harm. However, adversarial users can construct inputs which circumvent attempts at alignment. In this work, we study adversarial alignment, and ask to what extent these models remain aligned when interacting with an adversarial user who constructs worst-case inputs (adversarial examples). These inputs are designed to cause the model to emit harmful content that would otherwise be prohibited. We show that existing NLP-based optimization attacks are insufficiently powerful to reliably attack aligned text models: even when current NLP-based attacks fail, we can find adversarial inputs with brute force. As a result, the failure of current attacks should not be seen as proof that aligned text models remain aligned under adversarial inputs. However the recent trend in large-scale ML models is multimodal models that allow users to provide images that influence the text that is generated. We show these models can be easily attacked, i.e., induced to perform arbitrary un-aligned behavior through adversarial perturbation of the input image. We conjecture that improved NLP attacks may demonstrate this same level of adversarial control over text-only models.

Nicholas Carlini, Florian Tramer, Krishnamurthy Dj Dvijotham et al. · eth-zurich

In this paper we show how to achieve state-of-the-art certified adversarial robustness to 2-norm bounded perturbations by relying exclusively on off-the-shelf pretrained models. To do so, we instantiate the denoised smoothing approach of Salman et al. 2020 by combining a pretrained denoising diffusion probabilistic model and a standard high-accuracy classifier. This allows us to certify 71% accuracy on ImageNet under adversarial perturbations constrained to be within an 2-norm of 0.5, an improvement of 14 percentage points over the prior certified SoTA using any approach, or an improvement of 30 percentage points over denoised smoothing. We obtain these results using only pretrained diffusion models and image classifiers, without requiring any fine tuning or retraining of model parameters.

Nicholas Carlini, Matthew Jagielski, Chiyuan Zhang et al. · eth-zurich

Machine learning models trained on private datasets have been shown to leak their private data. While recent work has found that the average data point is rarely leaked, the outlier samples are frequently subject to memorization and, consequently, privacy leakage. We demonstrate and analyse an Onion Effect of memorization: removing the "layer" of outlier points that are most vulnerable to a privacy attack exposes a new layer of previously-safe points to the same attack. We perform several experiments to study this effect, and understand why it occurs. The existence of this effect has various consequences. For example, it suggests that proposals to defend against memorization without training with rigorous privacy guarantees are unlikely to be effective. Further, it suggests that privacy-enhancing technologies such as machine unlearning could actually harm the privacy of other users.

Roland S. Zimmermann, Wieland Brendel, Florian Tramer et al. · eth-zurich

Hundreds of defenses have been proposed to make deep neural networks robust against minimal (adversarial) input perturbations. However, only a handful of these defenses held up their claims because correctly evaluating robustness is extremely challenging: Weak attacks often fail to find adversarial examples even if they unknowingly exist, thereby making a vulnerable network look robust. In this paper, we propose a test to identify weak attacks, and thus weak defense evaluations. Our test slightly modifies a neural network to guarantee the existence of an adversarial example for every sample. Consequentially, any correct attack must succeed in breaking this modified network. For eleven out of thirteen previously-published defenses, the original evaluation of the defense fails our test, while stronger attacks that break these defenses pass it. We hope that attack unit tests - such as ours - will be a major component in future robustness evaluations and increase confidence in an empirical field that is currently riddled with skepticism.

Patrick Chao, Edoardo Debenedetti, Alexander Robey et al. · eth-zurich, princeton

Jailbreak attacks cause large language models (LLMs) to generate harmful, unethical, or otherwise objectionable content. Evaluating these attacks presents a number of challenges, which the current collection of benchmarks and evaluation techniques do not adequately address. First, there is no clear standard of practice regarding jailbreaking evaluation. Second, existing works compute costs and success rates in incomparable ways. And third, numerous works are not reproducible, as they withhold adversarial prompts, involve closed-source code, or rely on evolving proprietary APIs. To address these challenges, we introduce JailbreakBench, an open-sourced benchmark with the following components: (1) an evolving repository of state-of-the-art adversarial prompts, which we refer to as jailbreak artifacts; (2) a jailbreaking dataset comprising 100 behaviors -- both original and sourced from prior work (Zou et al., 2023; Mazeika et al., 2023, 2024) -- which align with OpenAI's usage policies; (3) a standardized evaluation framework at https://github.com/JailbreakBench/jailbreakbench that includes a clearly defined threat model, system prompts, chat templates, and scoring functions; and (4) a leaderboard at https://jailbreakbench.github.io/ that tracks the performance of attacks and defenses for various LLMs. We have carefully considered the potential ethical implications of releasing this benchmark, and believe that it will be a net positive for the community.

Xin Chen, Jie Zhang, Florian Tramer

Prompt injection is one of the most critical vulnerabilities in LLM agents; yet, effective automated attacks remain largely unexplored from an optimization perspective. Existing methods heavily depend on human red-teamers and hand-crafted prompts, limiting their scalability and adaptability. We propose AutoInject, a reinforcement learning framework that generates universal, transferable adversarial suffixes while jointly optimizing for attack success and utility preservation on benign tasks. Our black-box method supports both query-based optimization and transfer attacks to unseen models and tasks. Using only a 1.5B parameter adversarial suffix generator, we successfully compromise frontier systems including GPT 5 Nano, Claude Sonnet 3.5, and Gemini 2.5 Flash on the AgentDojo benchmark, establishing a stronger baseline for automated prompt injection research.

Usman Anwar, Abulhair Saparov, Javier Rando et al. · cambridge, eth-zurich

This work identifies 18 foundational challenges in assuring the alignment and safety of large language models (LLMs). These challenges are organized into three different categories: scientific understanding of LLMs, development and deployment methods, and sociotechnical challenges. Based on the identified challenges, we pose $200+$ concrete research questions.

Florian Tramer, Andreas Terzis, Thomas Steinke et al.

Differential Privacy can provide provable privacy guarantees for training data in machine learning. However, the presence of proofs does not preclude the presence of errors. Inspired by recent advances in auditing which have been used for estimating lower bounds on differentially private algorithms, here we show that auditing can also be used to find flaws in (purportedly) differentially private schemes. In this case study, we audit a recent open source implementation of a differentially private deep learning algorithm and find, with 99.99999999% confidence, that the implementation does not satisfy the claimed differential privacy guarantee.

Xuandong Zhao, Sam Gunn, Miranda Christ et al. · berkeley, eth-zurich

As the outputs of generative AI (GenAI) techniques improve in quality, it becomes increasingly challenging to distinguish them from human-created content. Watermarking schemes are a promising approach to address the problem of distinguishing between AI and human-generated content. These schemes embed hidden signals within AI-generated content to enable reliable detection. While watermarking is not a silver bullet for addressing all risks associated with GenAI, it can play a crucial role in enhancing AI safety and trustworthiness by combating misinformation and deception. This paper presents a comprehensive overview of watermarking techniques for GenAI, beginning with the need for watermarking from historical and regulatory perspectives. We formalize the definitions and desired properties of watermarking schemes and examine the key objectives and threat models for existing approaches. Practical evaluation strategies are also explored, providing insights into the development of robust watermarking techniques capable of resisting various attacks. Additionally, we review recent representative works, highlight open challenges, and discuss potential directions for this emerging field. By offering a thorough understanding of watermarking in GenAI, this work aims to guide researchers in advancing watermarking methods and applications, and support policymakers in addressing the broader implications of GenAI.

Yangsibo Huang, Milad Nasr, Anastasios Angelopoulos et al. · eth-zurich

It is now common to evaluate Large Language Models (LLMs) by having humans manually vote to evaluate model outputs, in contrast to typical benchmarks that evaluate knowledge or skill at some particular task. Chatbot Arena, the most popular benchmark of this type, ranks models by asking users to select the better response between two randomly selected models (without revealing which model was responsible for the generations). These platforms are widely trusted as a fair and accurate measure of LLM capabilities. In this paper, we show that if bot protection and other defenses are not implemented, these voting-based benchmarks are potentially vulnerable to adversarial manipulation. Specifically, we show that an attacker can alter the leaderboard (to promote their favorite model or demote competitors) at the cost of roughly a thousand votes (verified in a simulated, offline version of Chatbot Arena). Our attack consists of two steps: first, we show how an attacker can determine which model was used to generate a given reply with more than $95\%$ accuracy; and then, the attacker can use this information to consistently vote for (or against) a target model. Working with the Chatbot Arena developers, we identify, propose, and implement mitigations to improve the robustness of Chatbot Arena against adversarial manipulation, which, based on our analysis, substantially increases the cost of such attacks. Some of these defenses were present before our collaboration, such as bot protection with Cloudflare, malicious user detection, and rate limiting. Others, including reCAPTCHA and login are being integrated to strengthen the security in Chatbot Arena.

Nicholas Carlini, Daphne Ippolito, Matthew Jagielski et al.

Large language models (LMs) have been shown to memorize parts of their training data, and when prompted appropriately, they will emit the memorized training data verbatim. This is undesirable because memorization violates privacy (exposing user data), degrades utility (repeated easy-to-memorize text is often low quality), and hurts fairness (some texts are memorized over others). We describe three log-linear relationships that quantify the degree to which LMs emit memorized training data. Memorization significantly grows as we increase (1) the capacity of a model, (2) the number of times an example has been duplicated, and (3) the number of tokens of context used to prompt the model. Surprisingly, we find the situation becomes more complicated when generalizing these results across model families. On the whole, we find that memorization in LMs is more prevalent than previously believed and will likely get worse as models continues to scale, at least without active mitigations.

Nicholas Carlini, Steve Chien, Milad Nasr et al.

A membership inference attack allows an adversary to query a trained machine learning model to predict whether or not a particular example was contained in the model's training dataset. These attacks are currently evaluated using average-case "accuracy" metrics that fail to characterize whether the attack can confidently identify any members of the training set. We argue that attacks should instead be evaluated by computing their true-positive rate at low (e.g., <0.1%) false-positive rates, and find most prior attacks perform poorly when evaluated in this way. To address this we develop a Likelihood Ratio Attack (LiRA) that carefully combines multiple ideas from the literature. Our attack is 10x more powerful at low false-positive rates, and also strictly dominates prior attacks on existing metrics.

Nicholas Carlini, Sanjam Garg, Somesh Jha et al.

NeuraCrypt (Yara et al. arXiv 2021) is an algorithm that converts a sensitive dataset to an encoded dataset so that (1) it is still possible to train machine learning models on the encoded data, but (2) an adversary who has access only to the encoded dataset can not learn much about the original sensitive dataset. We break NeuraCrypt privacy claims, by perfectly solving the authors' public challenge, and by showing that NeuraCrypt does not satisfy the formal privacy definitions posed in the original paper. Our attack consists of a series of boosting steps that, coupled with various design flaws, turns a 1% attack advantage into a 100% complete break of the scheme.

Nicholas Carlini, Florian Tramer, Eric Wallace et al.

It has become common to publish large (billion parameter) language models that have been trained on private datasets. This paper demonstrates that in such settings, an adversary can perform a training data extraction attack to recover individual training examples by querying the language model. We demonstrate our attack on GPT-2, a language model trained on scrapes of the public Internet, and are able to extract hundreds of verbatim text sequences from the model's training data. These extracted examples include (public) personally identifiable information (names, phone numbers, and email addresses), IRC conversations, code, and 128-bit UUIDs. Our attack is possible even though each of the above sequences are included in just one document in the training data. We comprehensively evaluate our extraction attack to understand the factors that contribute to its success. Worryingly, we find that larger models are more vulnerable than smaller models. We conclude by drawing lessons and discussing possible safeguards for training large language models.

Nicholas Carlini, Samuel Deng, Sanjam Garg et al.

A private machine learning algorithm hides as much as possible about its training data while still preserving accuracy. In this work, we study whether a non-private learning algorithm can be made private by relying on an instance-encoding mechanism that modifies the training inputs before feeding them to a normal learner. We formalize both the notion of instance encoding and its privacy by providing two attack models. We first prove impossibility results for achieving a (stronger) model. Next, we demonstrate practical attacks in the second (weaker) attack model on InstaHide, a recent proposal by Huang, Song, Li and Arora [ICML'20] that aims to use instance encoding for privacy.

Christopher A. Choquette-Choo, Florian Tramer, Nicholas Carlini et al.

Membership inference attacks are one of the simplest forms of privacy leakage for machine learning models: given a data point and model, determine whether the point was used to train the model. Existing membership inference attacks exploit models' abnormal confidence when queried on their training data. These attacks do not apply if the adversary only gets access to models' predicted labels, without a confidence measure. In this paper, we introduce label-only membership inference attacks. Instead of relying on confidence scores, our attacks evaluate the robustness of a model's predicted labels under perturbations to obtain a fine-grained membership signal. These perturbations include common data augmentations or adversarial examples. We empirically show that our label-only membership inference attacks perform on par with prior attacks that required access to model confidences. We further demonstrate that label-only attacks break multiple defenses against membership inference attacks that (implicitly or explicitly) rely on a phenomenon we call confidence masking. These defenses modify a model's confidence scores in order to thwart attacks, but leave the model's predicted labels unchanged. Our label-only attacks demonstrate that confidence-masking is not a viable defense strategy against membership inference. Finally, we investigate worst-case label-only attacks, that infer membership for a small number of outlier data points. We show that label-only attacks also match confidence-based attacks in this setting. We find that training models with differential privacy and (strong) L2 regularization are the only known defense strategies that successfully prevents all attacks. This remains true even when the differential privacy budget is too high to offer meaningful provable guarantees.

Florian Tramer, Nicholas Carlini, Wieland Brendel et al.

Adaptive attacks have (rightfully) become the de facto standard for evaluating defenses to adversarial examples. We find, however, that typical adaptive evaluations are incomplete. We demonstrate that thirteen defenses recently published at ICLR, ICML and NeurIPS---and chosen for illustrative and pedagogical purposes---can be circumvented despite attempting to perform evaluations using adaptive attacks. While prior evaluation papers focused mainly on the end result---showing that a defense was ineffective---this paper focuses on laying out the methodology and the approach necessary to perform an adaptive attack. We hope that these analyses will serve as guidance on how to properly perform adaptive attacks against defenses to adversarial examples, and thus will allow the community to make further progress in building more robust models.

Charlie Hou, Mingxun Zhou, Yan Ji et al.

Incentive mechanisms are central to the functionality of permissionless blockchains: they incentivize participants to run and secure the underlying consensus protocol. Designing incentive-compatible incentive mechanisms is notoriously challenging, however. As a result, most public blockchains today use incentive mechanisms whose security properties are poorly understood and largely untested. In this work, we propose SquirRL, a framework for using deep reinforcement learning to analyze attacks on blockchain incentive mechanisms. We demonstrate SquirRL's power by first recovering known attacks: (1) the optimal selfish mining attack in Bitcoin [52], and (2) the Nash equilibrium in block withholding attacks [16]. We also use SquirRL to obtain several novel empirical results. First, we discover a counterintuitive flaw in the widely used rushing adversary model when applied to multi-agent Markov games with incomplete information. Second, we demonstrate that the optimal selfish mining strategy identified in [52] is actually not a Nash equilibrium in the multi-agent selfish mining setting. In fact, our results suggest (but do not prove) that when more than two competing agents engage in selfish mining, there is no profitable Nash equilibrium. This is consistent with the lack of observed selfish mining in the wild. Third, we find a novel attack on a simplified version of Ethereum's finalization mechanism, Casper the Friendly Finality Gadget (FFG) that allows a strategic agent to amplify her rewards by up to 30%. Notably, [10] show that honest voting is a Nash equilibrium in Casper FFG: our attack shows that when Casper FFG is composed with selfish mining, this is no longer the case. Altogether, our experiments demonstrate SquirRL's flexibility and promise as a framework for studying attack settings that have thus far eluded theoretical and empirical understanding.

Kevin Eykholt, Ivan Evtimov, Earlence Fernandes et al.

Deep neural networks (DNNs) are vulnerable to adversarial examples-maliciously crafted inputs that cause DNNs to make incorrect predictions. Recent work has shown that these attacks generalize to the physical domain, to create perturbations on physical objects that fool image classifiers under a variety of real-world conditions. Such attacks pose a risk to deep learning models used in safety-critical cyber-physical systems. In this work, we extend physical attacks to more challenging object detection models, a broader class of deep learning algorithms widely used to detect and label multiple objects within a scene. Improving upon a previous physical attack on image classifiers, we create perturbed physical objects that are either ignored or mislabeled by object detection models. We implement a Disappearance Attack, in which we cause a Stop sign to "disappear" according to the detector-either by covering thesign with an adversarial Stop sign poster, or by adding adversarial stickers onto the sign. In a video recorded in a controlled lab environment, the state-of-the-art YOLOv2 detector failed to recognize these adversarial Stop signs in over 85% of the video frames. In an outdoor experiment, YOLO was fooled by the poster and sticker attacks in 72.5% and 63.5% of the video frames respectively. We also use Faster R-CNN, a different object detection model, to demonstrate the transferability of our adversarial perturbations. The created poster perturbation is able to fool Faster R-CNN in 85.9% of the video frames in a controlled lab environment, and 40.2% of the video frames in an outdoor environment. Finally, we present preliminary results with a new Creation Attack, where in innocuous physical stickers fool a model into detecting nonexistent objects.

Kevin Eykholt, Ivan Evtimov, Earlence Fernandes et al.

Deep learning has proven to be a powerful tool for computer vision and has seen widespread adoption for numerous tasks. However, deep learning algorithms are known to be vulnerable to adversarial examples. These adversarial inputs are created such that, when provided to a deep learning algorithm, they are very likely to be mislabeled. This can be problematic when deep learning is used to assist in safety critical decisions. Recent research has shown that classifiers can be attacked by physical adversarial examples under various physical conditions. Given the fact that state-of-the-art objection detection algorithms are harder to be fooled by the same set of adversarial examples, here we show that these detectors can also be attacked by physical adversarial examples. In this note, we briefly show both static and dynamic test results. We design an algorithm that produces physical adversarial inputs, which can fool the YOLO object detector and can also attack Faster-RCNN with relatively high success rate based on transferability. Furthermore, our algorithm can compress the size of the adversarial inputs to stickers that, when attached to the targeted object, result in the detector either mislabeling or not detecting the object a high percentage of the time. This note provides a small set of results. Our upcoming paper will contain a thorough evaluation on other object detectors, and will present the algorithm.