Ahmed Salem

Nils Lukas, Ahmed Salem, Robert Sim et al.

Language Models (LMs) have been shown to leak information about training data through sentence-level membership inference and reconstruction attacks. Understanding the risk of LMs leaking Personally Identifiable Information (PII) has received less attention, which can be attributed to the false assumption that dataset curation techniques such as scrubbing are sufficient to prevent PII leakage. Scrubbing techniques reduce but do not prevent the risk of PII leakage: in practice scrubbing is imperfect and must balance the trade-off between minimizing disclosure and preserving the utility of the dataset. On the other hand, it is unclear to which extent algorithmic defenses such as differential privacy, designed to guarantee sentence- or user-level privacy, prevent PII disclosure. In this work, we introduce rigorous game-based definitions for three types of PII leakage via black-box extraction, inference, and reconstruction attacks with only API access to an LM. We empirically evaluate the attacks against GPT-2 models fine-tuned with and without defenses in three domains: case law, health care, and e-mails. Our main contributions are (i) novel attacks that can extract up to 10$\times$ more PII sequences than existing attacks, (ii) showing that sentence-level differential privacy reduces the risk of PII disclosure but still leaks about 3% of PII sequences, and (iii) a subtle connection between record-level membership inference and PII reconstruction. Code to reproduce all experiments in the paper is available at https://github.com/microsoft/analysing_pii_leakage.

Zheng Li, Ning Yu, Ahmed Salem et al.

Deepfakes pose severe threats of visual misinformation to our society. One representative deepfake application is face manipulation that modifies a victim's facial attributes in an image, e.g., changing her age or hair color. The state-of-the-art face manipulation techniques rely on Generative Adversarial Networks (GANs). In this paper, we propose the first defense system, namely UnGANable, against GAN-inversion-based face manipulation. In specific, UnGANable focuses on defending GAN inversion, an essential step for face manipulation. Its core technique is to search for alternative images (called cloaked images) around the original images (called target images) in image space. When posted online, these cloaked images can jeopardize the GAN inversion process. We consider two state-of-the-art inversion techniques including optimization-based inversion and hybrid inversion, and design five different defenses under five scenarios depending on the defender's background knowledge. Extensive experiments on four popular GAN models trained on two benchmark face datasets show that UnGANable achieves remarkable effectiveness and utility performance, and outperforms multiple baseline methods. We further investigate four adaptive adversaries to bypass UnGANable and show that some of them are slightly effective.

Ziqing Yang, Michael Backes, Yang Zhang et al.

Open-source large language models (LLMs) have become increasingly popular among both the general public and industry, as they can be customized, fine-tuned, and freely used. However, some open-source LLMs require approval before usage, which has led to third parties publishing their own easily accessible versions. Similarly, third parties have been publishing fine-tuned or quantized variants of these LLMs. These versions are particularly appealing to users because of their ease of access and reduced computational resource demands. This trend has increased the risk of training time attacks, compromising the integrity and security of LLMs. In this work, we present a new training time attack, SOS, which is designed to be low in computational demand and does not require clean data or modification of the model weights, thereby maintaining the model's utility intact. The attack addresses security issues in various scenarios, including the backdoor attack, jailbreak attack, and prompt stealing attack. Our experimental findings demonstrate that the proposed attack is effective across all evaluated targets. Furthermore, we present the other side of our SOS technique, namely the copyright token -- a novel technique that enables users to mark their copyrighted content and prevent models from using it.

Boyang Zhang, Yicong Tan, Yun Shen et al.

Recently, autonomous agents built on large language models (LLMs) have experienced significant development and are being deployed in real-world applications. These agents can extend the base LLM's capabilities in multiple ways. For example, a well-built agent using GPT-3.5-Turbo as its core can outperform the more advanced GPT-4 model by leveraging external components. More importantly, the usage of tools enables these systems to perform actions in the real world, moving from merely generating text to actively interacting with their environment. Given the agents' practical applications and their ability to execute consequential actions, it is crucial to assess potential vulnerabilities. Such autonomous systems can cause more severe damage than a standalone language model if compromised. While some existing research has explored harmful actions by LLM agents, our study approaches the vulnerability from a different perspective. We introduce a new type of attack that causes malfunctions by misleading the agent into executing repetitive or irrelevant actions. We conduct comprehensive evaluations using various attack methods, surfaces, and properties to pinpoint areas of susceptibility. Our experiments reveal that these attacks can induce failure rates exceeding 80\% in multiple scenarios. Through attacks on implemented and deployable agents in multi-agent scenarios, we accentuate the realistic risks associated with these vulnerabilities. To mitigate such attacks, we propose self-examination detection methods. However, our findings indicate these attacks are difficult to detect effectively using LLMs alone, highlighting the substantial risks associated with this vulnerability.

Ahmed Salem, Giovanni Cherubin, David Evans et al.

Deploying machine learning models in production may allow adversaries to infer sensitive information about training data. There is a vast literature analyzing different types of inference risks, ranging from membership inference to reconstruction attacks. Inspired by the success of games (i.e., probabilistic experiments) to study security properties in cryptography, some authors describe privacy inference risks in machine learning using a similar game-based style. However, adversary capabilities and goals are often stated in subtly different ways from one presentation to the other, which makes it hard to relate and compose results. In this paper, we present a game-based framework to systematize the body of knowledge on privacy inference risks in machine learning. We use this framework to (1) provide a unifying structure for definitions of inference risks, (2) formally establish known relations among definitions, and (3) to uncover hitherto unknown relations that would have been difficult to spot otherwise.

Mark Russinovich, Ahmed Salem

Growing concerns over the theft and misuse of Large Language Models (LLMs) have heightened the need for effective fingerprinting, which links a model to its original version to detect misuse. In this paper, we define five key properties for a successful fingerprint: Transparency, Efficiency, Persistence, Robustness, and Unforgeability. We introduce a novel fingerprinting framework that provides verifiable proof of ownership while maintaining fingerprint integrity. Our approach makes two main contributions. First, we propose a Chain and Hash technique that cryptographically binds fingerprint prompts with their responses, ensuring no adversary can generate colliding fingerprints and allowing model owners to irrefutably demonstrate their creation. Second, we address a realistic threat model in which instruction-tuned models' output distribution can be significantly altered through meta-prompts. By integrating random padding and varied meta-prompt configurations during training, our method preserves fingerprint robustness even when the model's output style is significantly modified. Experimental results demonstrate that our framework offers strong security for proving ownership and remains resilient against benign transformations like fine-tuning, as well as adversarial attempts to erase fingerprints. Finally, we also demonstrate its applicability to fingerprinting LoRA adapters.

Santiago Zanella-Béguelin, Lukas Wutschitz, Shruti Tople et al.

Algorithms such as Differentially Private SGD enable training machine learning models with formal privacy guarantees. However, there is a discrepancy between the protection that such algorithms guarantee in theory and the protection they afford in practice. An emerging strand of work empirically estimates the protection afforded by differentially private training as a confidence interval for the privacy budget $\varepsilon$ spent on training a model. Existing approaches derive confidence intervals for $\varepsilon$ from confidence intervals for the false positive and false negative rates of membership inference attacks. Unfortunately, obtaining narrow high-confidence intervals for $ε$ using this method requires an impractically large sample size and training as many models as samples. We propose a novel Bayesian method that greatly reduces sample size, and adapt and validate a heuristic to draw more than one sample per trained model. Our Bayesian method exploits the hypothesis testing interpretation of differential privacy to obtain a posterior for $\varepsilon$ (not just a confidence interval) from the joint posterior of the false positive and false negative rates of membership inference attacks. For the same sample size and confidence, we derive confidence intervals for $\varepsilon$ around 40% narrower than prior work. The heuristic, which we adapt from label-only DP, can be used to further reduce the number of trained models needed to get enough samples by up to 2 orders of magnitude.

Lukas Wutschitz, Boris Köpf, Andrew Paverd et al.

Modern machine learning systems use models trained on ever-growing corpora. Typically, metadata such as ownership, access control, or licensing information is ignored during training. Instead, to mitigate privacy risks, we rely on generic techniques such as dataset sanitization and differentially private model training, with inherent privacy/utility trade-offs that hurt model performance. Moreover, these techniques have limitations in scenarios where sensitive information is shared across multiple participants and fine-grained access control is required. By ignoring metadata, we therefore miss an opportunity to better address security, privacy, and confidentiality challenges. In this paper, we take an information flow control perspective to describe machine learning systems, which allows us to leverage metadata such as access control policies and define clear-cut privacy and confidentiality guarantees with interpretable information flows. Under this perspective, we contrast two different approaches to achieve user-level non-interference: 1) fine-tuning per-user models, and 2) retrieval augmented models that access user-specific datasets at inference time. We compare these two approaches to a trivially non-interfering zero-shot baseline using a public model and to a baseline that fine-tunes this model on the whole corpus. We evaluate trained models on two datasets of scientific articles and demonstrate that retrieval augmented architectures deliver the best utility, scalability, and flexibility while satisfying strict non-interference guarantees.

Rui Wen, Tianhao Wang, Michael Backes et al.

Large Language Models (LLMs) are powerful tools for natural language processing, enabling novel applications and user experiences. However, to achieve optimal performance, LLMs often require adaptation with private data, which poses privacy and security challenges. Several techniques have been proposed to adapt LLMs with private data, such as Low-Rank Adaptation (LoRA), Soft Prompt Tuning (SPT), and In-Context Learning (ICL), but their comparative privacy and security properties have not been systematically investigated. In this work, we fill this gap by evaluating the robustness of LoRA, SPT, and ICL against three types of well-established attacks: membership inference, which exposes data leakage (privacy); backdoor, which injects malicious behavior (security); and model stealing, which can violate intellectual property (privacy and security). Our results show that there is no silver bullet for privacy and security in LLM adaptation and each technique has different strengths and weaknesses.

Adel Elmahdy, Ahmed Salem

Natural language processing (NLP) models have become increasingly popular in real-world applications, such as text classification. However, they are vulnerable to privacy attacks, including data reconstruction attacks that aim to extract the data used to train the model. Most previous studies on data reconstruction attacks have focused on LLM, while classification models were assumed to be more secure. In this work, we propose a new targeted data reconstruction attack called the Mix And Match attack, which takes advantage of the fact that most classification models are based on LLM. The Mix And Match attack uses the base model of the target model to generate candidate tokens and then prunes them using the classification head. We extensively demonstrate the effectiveness of the attack using both random and organic canaries. This work highlights the importance of considering the privacy risks associated with data reconstruction attacks in classification models and offers insights into possible leakages.

Ahmed Salem, Andrew Paverd, Sahar Abdelnabi

Large language models (LLMs) are commonly treated as stateless: once an interaction ends, no information is assumed to persist unless it is explicitly stored and re-supplied. We challenge this assumption by introducing implicit memory-the ability of a model to carry state across otherwise independent interactions by encoding information in its own outputs and later recovering it when those outputs are reintroduced as input. This mechanism does not require any explicit memory module, yet it creates a persistent information channel across inference requests. As a concrete demonstration, we introduce a new class of temporal backdoors, which we call time bombs. Unlike conventional backdoors that activate on a single trigger input, time bombs activate only after a sequence of interactions satisfies hidden conditions accumulated via implicit memory. We show that such behavior can be induced today through straightforward prompting or fine-tuning. Beyond this case study, we analyze broader implications of implicit memory, including covert inter-agent communication, benchmark contamination, targeted manipulation, and training-data poisoning. Finally, we discuss detection challenges and outline directions for stress-testing and evaluation, with the goal of anticipating and controlling future developments. To promote future research, we release code and data at: https://github.com/microsoft/implicitMemory.

Boyang Zhang, Xinyue Shen, Wai Man Si et al.

Moderating offensive, hateful, and toxic language has always been an important but challenging topic in the domain of safe use in NLP. The emerging large language models (LLMs), such as ChatGPT, can potentially further accentuate this threat. Previous works have discovered that ChatGPT can generate toxic responses using carefully crafted inputs. However, limited research has been done to systematically examine when ChatGPT generates toxic responses. In this paper, we comprehensively evaluate the toxicity in ChatGPT by utilizing instruction-tuning datasets that closely align with real-world scenarios. Our results show that ChatGPT's toxicity varies based on different properties and settings of the prompts, including tasks, domains, length, and languages. Notably, prompts in creative writing tasks can be 2x more likely than others to elicit toxic responses. Prompting in German and Portuguese can also double the response toxicity. Additionally, we discover that certain deliberately toxic prompts, designed in earlier studies, no longer yield harmful responses. We hope our discoveries can guide model developers to better regulate these AI systems and the users to avoid undesirable outputs.

Maximilian Kreutner, Jens Rupprecht, Georg Ahnert et al.

We introduce QSTN, an open-source Python framework for systematically generating responses from questionnaire-style prompts to support in-silico surveys and annotation tasks with large language models (LLMs). QSTN enables robust evaluation of questionnaire presentation, prompt perturbations, and response generation methods. Our extensive evaluation ($>40 $ million survey responses) shows that question structure and response generation methods have a significant impact on the alignment of generated survey responses with human answers, and can be obtained for a fraction of the compute cost. In addition, we offer a no-code user interface that allows researchers to set up robust experiments with LLMs without coding knowledge. We hope that QSTN will support the reproducibility and reliability of LLM-based research in the future.

Mark Russinovich, Yanan Cai, Keegan Hines et al.

Safety alignment is only as robust as its weakest failure mode. Despite extensive work on safety post-training, it has been shown that models can be readily unaligned through post-deployment fine-tuning. However, these methods often require extensive data curation and degrade model utility. In this work, we extend the practical limits of unalignment by introducing GRP-Obliteration (GRP-Oblit), a method that uses Group Relative Policy Optimization (GRPO) to directly remove safety constraints from target models. We show that a single unlabeled prompt is sufficient to reliably unalign safety-aligned models while largely preserving their utility, and that GRP-Oblit achieves stronger unalignment on average than existing state-of-the-art techniques. Moreover, GRP-Oblit generalizes beyond language models and can also unalign diffusion-based image generation systems. We evaluate GRP-Oblit on six utility benchmarks and five safety benchmarks across fifteen 7-20B parameter models, spanning instruct and reasoning models, as well as dense and MoE architectures. The evaluated model families include GPT-OSS, distilled DeepSeek, Gemma, Llama, Ministral, and Qwen.

Minxing Zhang, Ahmed Salem, Michael Backes et al.

The increasing cost of training machine learning (ML) models has led to the inclusion of new parties to the training pipeline, such as users who contribute training data and companies that provide computing resources. This involvement of such new parties in the ML training process has introduced new attack surfaces for an adversary to exploit. A recent attack in this domain is the model hijacking attack, whereby an adversary hijacks a victim model to implement their own -- possibly malicious -- hijacking tasks. However, the scope of the model hijacking attack is so far limited to the homogeneous-modality tasks. In this paper, we transform the model hijacking attack into a more general multimodal setting, where the hijacking and original tasks are performed on data of different modalities. Specifically, we focus on the setting where an adversary implements a natural language processing (NLP) hijacking task into an image classification model. To mount the attack, we propose a novel encoder-decoder based framework, namely the Blender, which relies on advanced image and language models. Experimental results show that our modal hijacking attack achieves strong performances in different settings. For instance, our attack achieves 94%, 94%, and 95% attack success rate when using the Sogou news dataset to hijack STL10, CIFAR-10, and MNIST classifiers.

Manuel Costa, Boris Köpf, Aashish Kolluri et al.

As AI agents become increasingly autonomous and capable, ensuring their security against vulnerabilities such as prompt injection becomes critical. This paper explores the use of information-flow control (IFC) to provide security guarantees for AI agents. We present a formal model to reason about the security and expressiveness of agent planners. Using this model, we characterize the class of properties enforceable by dynamic taint-tracking and construct a taxonomy of tasks to evaluate security and utility trade-offs of planner designs. Informed by this exploration, we present Fides, a planner that tracks confidentiality and integrity labels, deterministically enforces security policies, and introduces novel primitives for selectively hiding information. Its evaluation in AgentDojo demonstrates that this approach enables us to complete a broad range of tasks with security guarantees. A tutorial to walk readers through the the concepts introduced in the paper can be found at https://github.com/microsoft/fides

Rui Wen, Mark Russinovich, Andrew Paverd et al.

Backdoor attacks pose a serious security threat to large language models (LLMs), which are increasingly deployed as general-purpose assistants in safety- and privacy-critical applications. Existing LLM backdoors rely primarily on content-based triggers, requiring explicit modification of the input text. In this work, we show that this assumption is unnecessary and limiting. We introduce MetaBackdoor, a new class of backdoor attacks that exploits positional information as the trigger, without modifying textual content. Our key insight is that Transformer-based LLMs necessarily encode token positions to process ordered sequences. As a result, length-correlated positional structure is reflected in the model's internal computation and can be used as an effective non-content trigger signal. We demonstrate that even a simple length-based positional trigger is sufficient to activate stealthy backdoors. Unlike prior attacks, MetaBackdoor operates on visibly and semantically clean inputs and enables qualitatively new capabilities. We show that a backdoored LLM can be induced to disclose sensitive internal information, including proprietary system prompts, once a length condition is satisfied. We further demonstrate a self-activation scenario, where normal multi-turn interaction can move the conversation context into the trigger region and induce malicious tool-call behavior without attacker-supplied trigger text. In addition, MetaBackdoor is orthogonal to content-based backdoors and can be composed with them to create more precise and harder-to-detect activation conditions. Our results expand the threat model of LLM backdoors by revealing positional encoding as a previously overlooked attack surface. This challenges defenses that focus on detecting suspicious text and highlights the need for new defense strategies that explicitly account for positional triggers in modern LLM architectures.

Mark Russinovich, Ahmed Salem

We introduce the Context Compliance Attack (CCA), a novel, optimization-free method for bypassing AI safety mechanisms. Unlike current approaches -- which rely on complex prompt engineering and computationally intensive optimization -- CCA exploits a fundamental architectural vulnerability inherent in many deployed AI systems. By subtly manipulating conversation history, CCA convinces the model to comply with a fabricated dialogue context, thereby triggering restricted behavior. Our evaluation across a diverse set of open-source and proprietary models demonstrates that this simple attack can circumvent state-of-the-art safety protocols. We discuss the implications of these findings and propose practical mitigation strategies to fortify AI systems against such elementary yet effective adversarial tactics.

Mark Russinovich, Ahmed Salem, Ronen Eldan

Large Language Models (LLMs) have risen significantly in popularity and are increasingly being adopted across multiple applications. These LLMs are heavily aligned to resist engaging in illegal or unethical topics as a means to avoid contributing to responsible AI harms. However, a recent line of attacks, known as jailbreaks, seek to overcome this alignment. Intuitively, jailbreak attacks aim to narrow the gap between what the model can do and what it is willing to do. In this paper, we introduce a novel jailbreak attack called Crescendo. Unlike existing jailbreak methods, Crescendo is a simple multi-turn jailbreak that interacts with the model in a seemingly benign manner. It begins with a general prompt or question about the task at hand and then gradually escalates the dialogue by referencing the model's replies progressively leading to a successful jailbreak. We evaluate Crescendo on various public systems, including ChatGPT, Gemini Pro, Gemini-Ultra, LlaMA-2 70b and LlaMA-3 70b Chat, and Anthropic Chat. Our results demonstrate the strong efficacy of Crescendo, with it achieving high attack success rates across all evaluated models and tasks. Furthermore, we present Crescendomation, a tool that automates the Crescendo attack and demonstrate its efficacy against state-of-the-art models through our evaluations. Crescendomation surpasses other state-of-the-art jailbreaking techniques on the AdvBench subset dataset, achieving 29-61% higher performance on GPT-4 and 49-71% on Gemini-Pro. Finally, we also demonstrate Crescendo's ability to jailbreak multimodal models.

Hatem Ibrahem, Ahmed Salem, Qinmin Vivian Hu et al.

Accurately estimating the 3D layout of rooms is a crucial task in computer vision, with potential applications in robotics, augmented reality, and interior design. This paper proposes a novel model, PanoTPS-Net, to estimate room layout from a single panorama image. Leveraging a Convolutional Neural Network (CNN) and incorporating a Thin Plate Spline (TPS) spatial transformation, the architecture of PanoTPS-Net is divided into two stages: First, a convolutional neural network extracts the high-level features from the input images, allowing the network to learn the spatial parameters of the TPS transformation. Second, the TPS spatial transformation layer is generated to warp a reference layout to the required layout based on the predicted parameters. This unique combination empowers the model to properly predict room layouts while also generalizing effectively to both cuboid and non-cuboid layouts. Extensive experiments on publicly available datasets and comparisons with state-of-the-art methods demonstrate the effectiveness of the proposed method. The results underscore the model's accuracy in room layout estimation and emphasize the compatibility between the TPS transformation and panorama images. The robustness of the model in handling both cuboid and non-cuboid room layout estimation is evident with a 3DIoU value of 85.49, 86.16, 81.76, and 91.98 on PanoContext, Stanford-2D3D, Matterport3DLayout, and ZInD datasets, respectively. The source code is available at: https://github.com/HatemHosam/PanoTPS_Net.

Edoardo Debenedetti, Javier Rando, Daniel Paleka et al.

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.

Amr Gomaa, Ahmed Salem, Sahar Abdelnabi

As language models evolve into autonomous agents that act and communicate on behalf of users, ensuring safety in multi-agent ecosystems becomes a central challenge. Interactions between personal assistants and external service providers expose a core tension between utility and protection: effective collaboration requires information sharing, yet every exchange creates new attack surfaces. We introduce ConVerse, a dynamic benchmark for evaluating privacy and security risks in agent-agent interactions. ConVerse spans three practical domains (travel, real estate, insurance) with 12 user personas and over 864 contextually grounded attacks (611 privacy, 253 security). Unlike prior single-agent settings, it models autonomous, multi-turn agent-to-agent conversations where malicious requests are embedded within plausible discourse. Privacy is tested through a three-tier taxonomy assessing abstraction quality, while security attacks target tool use and preference manipulation. Evaluating seven state-of-the-art models reveals persistent vulnerabilities; privacy attacks succeed in up to 88% of cases and security breaches in up to 60%, with stronger models leaking more. By unifying privacy and security within interactive multi-agent contexts, ConVerse reframes safety as an emergent property of communication.

Ahmed Salem, Andrew Paverd, Boris Köpf

Prompt injection has emerged as a serious security threat to large language models (LLMs). At present, the current best-practice for defending against newly-discovered prompt injection techniques is to add additional guardrails to the system (e.g., by updating the system prompt or using classifiers on the input and/or output of the model.) However, in the same way that variants of a piece of malware are created to evade anti-virus software, variants of a prompt injection can be created to evade the LLM's guardrails. Ideally, when a new prompt injection technique is discovered, candidate defenses should be tested not only against the successful prompt injection, but also against possible variants. In this work, we present, a tool to assist defenders in performing automated variant analysis of known prompt injection attacks. This involves solving two main challenges: (1) automatically generating variants of a given prompt according, and (2) automatically determining whether a variant was effective based only on the output of the model. This tool can also assist in generating datasets for jailbreak and prompt injection attacks, thus overcoming the scarcity of data in this domain. We evaluate Maatphor on three different types of prompt injection tasks. Starting from an ineffective (0%) seed prompt, Maatphor consistently generates variants that are at least 60% effective within the first 40 iterations.

Sahar Abdelnabi, Ahmed Salem

Reasoning-focused LLMs sometimes alter their behavior when they detect that they are being evaluated, which can lead them to optimize for test-passing performance or to comply more readily with harmful prompts if real-world consequences appear absent. We present the first quantitative study of how such "test awareness" impacts model behavior, particularly its performance on safety-related tasks. We introduce a white-box probing framework that (i) linearly identifies awareness-related activations and (ii) steers models toward or away from test awareness while monitoring downstream performance. We apply our method to different state-of-the-art open-weight reasoning LLMs across both realistic and hypothetical tasks (denoting tests or simulations). Our results demonstrate that test awareness significantly impacts safety alignment (such as compliance with harmful requests and conforming to stereotypes) with effects varying in both magnitude and direction across models. By providing control over this latent effect, our work aims to provide a stress-test mechanism and increase trust in how we perform safety evaluations.

Mark Russinovich, Ahmed Salem

Recent copyright agreements between AI companies and content creators underscore the need for fine-grained control over language models' ability to reproduce copyrighted text. Existing defenses-ranging from aggressive unlearning to simplistic output filters-either sacrifice model utility or inadequately address verbatim leakage. We introduce Obliviate, a lightweight post-training method that surgically suppresses exact reproduction of specified sequences while preserving semantic understanding. Obliviate first identifies memorized passages and then, for each target token, minimally adjusts the model's output distribution via a Kullback-Leibler divergence penalty to drive down the probability of exact reproduction. Simultaneously, we enforce a consistency loss on non-target tokens to retain the model's fluency and task performance. We evaluate Obliviate on four popular 6-8B-parameter models (LLaMA-3.1, LLaMA-3.1-Instruct, Qwen-2.5, and Yi-1.5) using synthetic memorization benchmarks and organic copyrighted excerpts (e.g., Moby Dick, Frankenstein, Alice in Wonderland and Les Miserables). Across all settings, Obliviate reduces verbatim recall by two orders of magnitude (e.g., from hundreds of words to fewer than 12) while degrading downstream accuracy by at most 1% on HellaSwag, MMLU, TruthfulQA, and Winogrande. Furthermore, we benchmark Obliviate aganist different unlearning and copyright techniques using the MUSE and CoTaEval benchmarks. These results position Obliviate as a practical, high-fidelity solution for copyright compliance in deployed LLMs.

Sahar Abdelnabi, Aideen Fay, Ahmed Salem et al.

Indirect Prompt Injection attacks exploit the inherent limitation of Large Language Models (LLMs) to distinguish between instructions and data in their inputs. Despite numerous defense proposals, the systematic evaluation against adaptive adversaries remains limited, even when successful attacks can have wide security and privacy implications, and many real-world LLM-based applications remain vulnerable. We present the results of LLMail-Inject, a public challenge simulating a realistic scenario in which participants adaptively attempted to inject malicious instructions into emails in order to trigger unauthorized tool calls in an LLM-based email assistant. The challenge spanned multiple defense strategies, LLM architectures, and retrieval configurations, resulting in a dataset of 208,095 unique attack submissions from 839 participants. We release the challenge code, the full dataset of submissions, and our analysis demonstrating how this data can provide new insights into the instruction-data separation problem. We hope this will serve as a foundation for future research towards practical structural solutions to prompt injection.

Blake Bullwinkel, Mark Russinovich, Ahmed Salem et al.

Recent research has demonstrated that state-of-the-art LLMs and defenses remain susceptible to multi-turn jailbreak attacks. These attacks require only closed-box model access and are often easy to perform manually, posing a significant threat to the safe and secure deployment of LLM-based systems. We study the effectiveness of the Crescendo multi-turn jailbreak at the level of intermediate model representations and find that safety-aligned LMs often represent Crescendo responses as more benign than harmful, especially as the number of conversation turns increases. Our analysis indicates that at each turn, Crescendo prompts tend to keep model outputs in a "benign" region of representation space, effectively tricking the model into fulfilling harmful requests. Further, our results help explain why single-turn jailbreak defenses like circuit breakers are generally ineffective against multi-turn attacks, motivating the development of mitigations that address this generalization gap.

Yanan Cai, Ahmed Salem, Besmira Nushi et al.

We introduce LogiPlan, a novel benchmark designed to evaluate the capabilities of large language models (LLMs) in logical planning and reasoning over complex relational structures. Logical relational reasoning is important for applications that may rely on LLMs to generate and query structured graphs of relations such as network infrastructure, knowledge bases, or business process schema. Our framework allows for dynamic variation of task complexity by controlling the number of objects, relations, and the minimum depth of relational chains, providing a fine-grained assessment of model performance across difficulty levels. LogiPlan encompasses three complementary tasks: (1) Plan Generation, where models must construct valid directed relational graphs meeting specified structural constraints; (2) Consistency Detection, testing models' ability to identify inconsistencies in relational structures; and (3) Comparison Question, evaluating models' capacity to determine the validity of queried relationships within a given graph. Additionally, we assess models' self-correction capabilities by prompting them to verify and refine their initial solutions. We evaluate state-of-the-art models including DeepSeek R1, Gemini 2.0 Pro, Gemini 2 Flash Thinking, GPT-4.5, GPT-4o, Llama 3.1 405B, O3-mini, O1, and Claude 3.7 Sonnet across these tasks, revealing significant performance gaps that correlate with model scale and architecture. Our analysis demonstrates that while recent reasoning-enhanced models show promising results on simpler instances, they struggle with more complex configurations requiring deeper logical planning.

Sahar Abdelnabi, Aideen Fay, Giovanni Cherubin et al.

LLMs are commonly used in retrieval-augmented applications to execute user instructions based on data from external sources. For example, modern search engines use LLMs to answer queries based on relevant search results; email plugins summarize emails by processing their content through an LLM. However, the potentially untrusted provenance of these data sources can lead to prompt injection attacks, where the LLM is manipulated by natural language instructions embedded in the external data, causing it to deviate from the user's original instruction(s). We define this deviation as task drift. Task drift is a significant concern as it allows attackers to exfiltrate data or influence the LLM's output for other users. We study LLM activations as a solution to detect task drift, showing that activation deltas - the difference in activations before and after processing external data - are strongly correlated with this phenomenon. Through two probing methods, we demonstrate that a simple linear classifier can detect drift with near-perfect ROC AUC on an out-of-distribution test set. We evaluate these methods by making minimal assumptions about how users' tasks, system prompts, and attacks can be phrased. We observe that this approach generalizes surprisingly well to unseen task domains, such as prompt injections, jailbreaks, and malicious instructions, without being trained on any of these attacks. Interestingly, the fact that this solution does not require any modifications to the LLM (e.g., fine-tuning), as well as its compatibility with existing meta-prompting solutions, makes it cost-efficient and easy to deploy. To encourage further research on activation-based task inspection, decoding, and interpretability, we release our large-scale TaskTracker toolkit, featuring a dataset of over 500K instances, representations from six SoTA language models, and a suite of inspection tools.

Wai Man Si, Michael Backes, Yang Zhang et al.

Machine learning has progressed significantly in various applications ranging from face recognition to text generation. However, its success has been accompanied by different attacks. Recently a new attack has been proposed which raises both accountability and parasitic computing risks, namely the model hijacking attack. Nevertheless, this attack has only focused on image classification tasks. In this work, we broaden the scope of this attack to include text generation and classification models, hence showing its broader applicability. More concretely, we propose a new model hijacking attack, Ditto, that can hijack different text classification tasks into multiple generation ones, e.g., language translation, text summarization, and language modeling. We use a range of text benchmark datasets such as SST-2, TweetEval, AGnews, QNLI, and IMDB to evaluate the performance of our attacks. Our results show that by using Ditto, an adversary can successfully hijack text generation models without jeopardizing their utility.

Ahmed Salem, Michael Backes, Yang Zhang

Machine learning (ML) has established itself as a cornerstone for various critical applications ranging from autonomous driving to authentication systems. However, with this increasing adoption rate of machine learning models, multiple attacks have emerged. One class of such attacks is training time attack, whereby an adversary executes their attack before or during the machine learning model training. In this work, we propose a new training time attack against computer vision based machine learning models, namely model hijacking attack. The adversary aims to hijack a target model to execute a different task than its original one without the model owner noticing. Model hijacking can cause accountability and security risks since a hijacked model owner can be framed for having their model offering illegal or unethical services. Model hijacking attacks are launched in the same way as existing data poisoning attacks. However, one requirement of the model hijacking attack is to be stealthy, i.e., the data samples used to hijack the target model should look similar to the model's original training dataset. To this end, we propose two different model hijacking attacks, namely Chameleon and Adverse Chameleon, based on a novel encoder-decoder style ML model, namely the Camouflager. Our evaluation shows that both of our model hijacking attacks achieve a high attack success rate, with a negligible drop in model utility.

Yugeng Liu, Rui Wen, Xinlei He et al.

Inference attacks against Machine Learning (ML) models allow adversaries to learn sensitive information about training data, model parameters, etc. While researchers have studied, in depth, several kinds of attacks, they have done so in isolation. As a result, we lack a comprehensive picture of the risks caused by the attacks, e.g., the different scenarios they can be applied to, the common factors that influence their performance, the relationship among them, or the effectiveness of possible defenses. In this paper, we fill this gap by presenting a first-of-its-kind holistic risk assessment of different inference attacks against machine learning models. We concentrate on four attacks -- namely, membership inference, model inversion, attribute inference, and model stealing -- and establish a threat model taxonomy. Our extensive experimental evaluation, run on five model architectures and four image datasets, shows that the complexity of the training dataset plays an important role with respect to the attack's performance, while the effectiveness of model stealing and membership inference attacks are negatively correlated. We also show that defenses like DP-SGD and Knowledge Distillation can only mitigate some of the inference attacks. Our analysis relies on a modular re-usable software, ML-Doctor, which enables ML model owners to assess the risks of deploying their models, and equally serves as a benchmark tool for researchers and practitioners.

Ahmed Salem, Michael Backes, Yang Zhang

Backdoor attack against deep neural networks is currently being profoundly investigated due to its severe security consequences. Current state-of-the-art backdoor attacks require the adversary to modify the input, usually by adding a trigger to it, for the target model to activate the backdoor. This added trigger not only increases the difficulty of launching the backdoor attack in the physical world, but also can be easily detected by multiple defense mechanisms. In this paper, we present the first triggerless backdoor attack against deep neural networks, where the adversary does not need to modify the input for triggering the backdoor. Our attack is based on the dropout technique. Concretely, we associate a set of target neurons that are dropped out during model training with the target label. In the prediction phase, the model will output the target label when the target neurons are dropped again, i.e., the backdoor attack is launched. This triggerless feature of our attack makes it practical in the physical world. Extensive experiments show that our triggerless backdoor attack achieves a perfect attack success rate with a negligible damage to the model's utility.

Ahmed Salem, Yannick Sautter, Michael Backes et al.

The tremendous progress of autoencoders and generative adversarial networks (GANs) has led to their application to multiple critical tasks, such as fraud detection and sanitized data generation. This increasing adoption has fostered the study of security and privacy risks stemming from these models. However, previous works have mainly focused on membership inference attacks. In this work, we explore one of the most severe attacks against machine learning models, namely the backdoor attack, against both autoencoders and GANs. The backdoor attack is a training time attack where the adversary implements a hidden backdoor in the target model that can only be activated by a secret trigger. State-of-the-art backdoor attacks focus on classification-based tasks. We extend the applicability of backdoor attacks to autoencoders and GAN-based models. More concretely, we propose the first backdoor attack against autoencoders and GANs where the adversary can control what the decoded or generated images are when the backdoor is activated. Our results show that the adversary can build a backdoored autoencoder that returns a target output for all backdoored inputs, while behaving perfectly normal on clean inputs. Similarly, for the GANs, our experiments show that the adversary can generate data from a different distribution when the backdoor is activated, while maintaining the same utility when the backdoor is not.

Xiaoyi Chen, Ahmed Salem, Dingfan Chen et al.

Deep neural networks (DNNs) have progressed rapidly during the past decade and have been deployed in various real-world applications. Meanwhile, DNN models have been shown to be vulnerable to security and privacy attacks. One such attack that has attracted a great deal of attention recently is the backdoor attack. Specifically, the adversary poisons the target model's training set to mislead any input with an added secret trigger to a target class. Previous backdoor attacks predominantly focus on computer vision (CV) applications, such as image classification. In this paper, we perform a systematic investigation of backdoor attack on NLP models, and propose BadNL, a general NLP backdoor attack framework including novel attack methods. Specifically, we propose three methods to construct triggers, namely BadChar, BadWord, and BadSentence, including basic and semantic-preserving variants. Our attacks achieve an almost perfect attack success rate with a negligible effect on the original model's utility. For instance, using the BadChar, our backdoor attack achieves a 98.9% attack success rate with yielding a utility improvement of 1.5% on the SST-5 dataset when only poisoning 3% of the original set. Moreover, we conduct a user study to prove that our triggers can well preserve the semantics from humans perspective.

Ahmed Salem, Rui Wen, Michael Backes et al.

Machine learning (ML) has made tremendous progress during the past decade and is being adopted in various critical real-world applications. However, recent research has shown that ML models are vulnerable to multiple security and privacy attacks. In particular, backdoor attacks against ML models have recently raised a lot of awareness. A successful backdoor attack can cause severe consequences, such as allowing an adversary to bypass critical authentication systems. Current backdooring techniques rely on adding static triggers (with fixed patterns and locations) on ML model inputs which are prone to detection by the current backdoor detection mechanisms. In this paper, we propose the first class of dynamic backdooring techniques against deep neural networks (DNN), namely Random Backdoor, Backdoor Generating Network (BaN), and conditional Backdoor Generating Network (c-BaN). Triggers generated by our techniques can have random patterns and locations, which reduce the efficacy of the current backdoor detection mechanisms. In particular, BaN and c-BaN based on a novel generative network are the first two schemes that algorithmically generate triggers. Moreover, c-BaN is the first conditional backdooring technique that given a target label, it can generate a target-specific trigger. Both BaN and c-BaN are essentially a general framework which renders the adversary the flexibility for further customizing backdoor attacks. We extensively evaluate our techniques on three benchmark datasets: MNIST, CelebA, and CIFAR-10. Our techniques achieve almost perfect attack performance on backdoored data with a negligible utility loss. We further show that our techniques can bypass current state-of-the-art defense mechanisms against backdoor attacks, including ABS, Februus, MNTD, Neural Cleanse, and STRIP.

Jinyuan Jia, Ahmed Salem, Michael Backes et al.

In a membership inference attack, an attacker aims to infer whether a data sample is in a target classifier's training dataset or not. Specifically, given a black-box access to the target classifier, the attacker trains a binary classifier, which takes a data sample's confidence score vector predicted by the target classifier as an input and predicts the data sample to be a member or non-member of the target classifier's training dataset. Membership inference attacks pose severe privacy and security threats to the training dataset. Most existing defenses leverage differential privacy when training the target classifier or regularize the training process of the target classifier. These defenses suffer from two key limitations: 1) they do not have formal utility-loss guarantees of the confidence score vectors, and 2) they achieve suboptimal privacy-utility tradeoffs. In this work, we propose MemGuard, the first defense with formal utility-loss guarantees against black-box membership inference attacks. Instead of tampering the training process of the target classifier, MemGuard adds noise to each confidence score vector predicted by the target classifier. Our key observation is that attacker uses a classifier to predict member or non-member and classifier is vulnerable to adversarial examples. Based on the observation, we propose to add a carefully crafted noise vector to a confidence score vector to turn it into an adversarial example that misleads the attacker's classifier. Our experimental results on three datasets show that MemGuard can effectively defend against membership inference attacks and achieve better privacy-utility tradeoffs than existing defenses. Our work is the first one to show that adversarial examples can be used as defensive mechanisms to defend against membership inference attacks.

Ahmed Salem, Apratim Bhattacharya, Michael Backes et al.

Machine learning (ML) has progressed rapidly during the past decade and the major factor that drives such development is the unprecedented large-scale data. As data generation is a continuous process, this leads to ML model owners updating their models frequently with newly-collected data in an online learning scenario. In consequence, if an ML model is queried with the same set of data samples at two different points in time, it will provide different results. In this paper, we investigate whether the change in the output of a black-box ML model before and after being updated can leak information of the dataset used to perform the update, namely the updating set. This constitutes a new attack surface against black-box ML models and such information leakage may compromise the intellectual property and data privacy of the ML model owner. We propose four attacks following an encoder-decoder formulation, which allows inferring diverse information of the updating set. Our new attacks are facilitated by state-of-the-art deep learning techniques. In particular, we propose a hybrid generative model (CBM-GAN) that is based on generative adversarial networks (GANs) but includes a reconstructive loss that allows reconstructing accurate samples. Our experiments show that the proposed attacks achieve strong performance.

Lucjan Hanzlik, Yang Zhang, Kathrin Grosse et al.

With the widespread use of machine learning (ML) techniques, ML as a service has become increasingly popular. In this setting, an ML model resides on a server and users can query it with their data via an API. However, if the user's input is sensitive, sending it to the server is undesirable and sometimes even legally not possible. Equally, the service provider does not want to share the model by sending it to the client for protecting its intellectual property and pay-per-query business model. In this paper, we propose MLCapsule, a guarded offline deployment of machine learning as a service. MLCapsule executes the model locally on the user's side and therefore the data never leaves the client. Meanwhile, MLCapsule offers the service provider the same level of control and security of its model as the commonly used server-side execution. In addition, MLCapsule is applicable to offline applications that require local execution. Beyond protecting against direct model access, we couple the secure offline deployment with defenses against advanced attacks on machine learning models such as model stealing, reverse engineering, and membership inference.

Ahmed Salem, Yang Zhang, Mathias Humbert et al.

Machine learning (ML) has become a core component of many real-world applications and training data is a key factor that drives current progress. This huge success has led Internet companies to deploy machine learning as a service (MLaaS). Recently, the first membership inference attack has shown that extraction of information on the training set is possible in such MLaaS settings, which has severe security and privacy implications. However, the early demonstrations of the feasibility of such attacks have many assumptions on the adversary, such as using multiple so-called shadow models, knowledge of the target model structure, and having a dataset from the same distribution as the target model's training data. We relax all these key assumptions, thereby showing that such attacks are very broadly applicable at low cost and thereby pose a more severe risk than previously thought. We present the most comprehensive study so far on this emerging and developing threat using eight diverse datasets which show the viability of the proposed attacks across domains. In addition, we propose the first effective defense mechanisms against such broader class of membership inference attacks that maintain a high level of utility of the ML model.