Retrieval-augmented generation (RAG) systems are vulnerable to attacks that inject poisoned passages into the retrieved context, even at low corruption rates. We show that existing attacks are not designed to be stealthy, allowing reliable detection and mitigation. We formalize a distinguishability-based security game to quantify stealth for such attacks. If a few poisoned passages control the response, they must bias the inference process more than the benign ones, inherently compromising stealth. This motivates analyzing intermediate signals of LLMs, such as attention weights, to approximate the influence of different passages on the response. Leveraging attention weights, we introduce the Normalized Passage Attention Score (NPAS) and a lightweight Attention-Variance Filter (AV Filter) that flags anomalous passages. Our method improves robustness, yielding up to ~20% higher accuracy than baseline defenses. We also develop adaptive attacks that attempt to conceal such anomalies, achieving up to 35% success rate and underscoring the challenges of achieving true stealth in poisoning RAG systems.

Leading language model (LM) providers like OpenAI and Anthopic allow customers to fine-tune frontier LMs for specific use cases. To prevent abuse, these providers apply filters to block fine-tuning on overtly harmful data. In this setting, we make three contributions: First, while past work has shown that safety alignment is superficial, we correspondingly demonstrate that existing fine-tuning attacks are "shallow" -- attacks target only the first several tokens of the model response, and consequently can be blocked by generating the first several response tokens with an aligned model. Second, we conceptually illustrate how to make attacks deeper by introducing a new fine-tuning attack that trains models to first refuse harmful requests before answering them; this ``refuse-then-comply" strategy bypasses shallow defenses and produces harmful responses that evade output filters. Third, we demonstrate the potency of our new fine-tuning attack by jailbreaking both open-source models equipped with defenses and production models, achieving attack success rates of 57% and 72% against GPT-4o and Claude Haiku, respectively. Our attack received a $2000 bug bounty from OpenAI and was acknowledged as a vulnerability by Anthropic.

AI agents equipped with tool-calling capabilities are susceptible to (IPI) attacks. In this attack scenario, malicious commands hidden within content trick the agent into performing unauthorized actions. Existing defenses can reduce attack success but often suffer from the : they deploy expensive, sanitization regardless of actual threat, thereby degrading utility and latency even in benign scenarios. We revisit IPI through a causal ablation perspective: a successful injection manifests as a where the user request no longer provides decisive support for the agent's privileged action, while a particular untrusted segment, such as a retrieved document or tool output, provides disproportionate attributable influence. Based on this signature, we propose , a selective defense framework that (i) computes lightweight, leave-one-out ablation-based attributions at privileged decision points, and (ii) triggers targeted sanitization only when an untrusted segment dominates the user intent. Additionally, CausalArmor employs to prevent the agent from acting on ``poisoned" reasoning traces. We present a theoretical analysis showing that sanitization based on attribution margins conditionally yields an exponentially small upper bound on the probability of selecting malicious actions. Experiments on AgentDojo and DoomArena demonstrate that CausalArmor matches the security of aggressive defenses while improving explainability and preserving utility and latency of AI agents.

Securing AI agents that operate in complex digital environments has become a critical need, and runtime monitoring approaches that formulate and enforce policies expressed in a formal language like Datalog offer a promising solution. However, existing approaches are restricted to deterministic policies. In many practical applications of AI agents, there is a need to enforce security policies in the face of ambiguity, leading to probabilistic predicates or state transitions (for example, a declassifier or Personally Identifiable Information (PII) detector that has some failure probability on each invocation). Furthermore, in many such applications, one cannot easily make the independence assumptions necessary to invoke prior work on probabilistic inference in Datalog. We address this by introducing a sound and efficient framework for such verification based on distributionally robust optimization, computing sound upper bounds on the probability of policy violation regardless of possible correlations between predicates. On standard benchmarks for terminal and tool calling agents, we demonstrate that our approach outperforms prior art and improves the security-utility trade-off while ensuring rigorous bounds on the probability of policy violation.

The rise of foundation models fine-tuned on human feedback from potentially untrusted users has increased the risk of adversarial data poisoning, necessitating the study of robustness of learning algorithms against such attacks. Existing research on provable certified robustness against data poisoning attacks primarily focuses on certifying robustness for static adversaries who modify a fraction of the dataset used to train the model before the training algorithm is applied. In practice, particularly when learning from human feedback in an online sense, adversaries can observe and react to the learning process and inject poisoned samples that optimize adversarial objectives better than when they are restricted to poisoning a static dataset once, before the learning algorithm is applied. Indeed, it has been shown in prior work that online dynamic adversaries can be significantly more powerful than static ones. We present a novel framework for computing certified bounds on the impact of dynamic poisoning, and use these certificates to design robust learning algorithms. We give an illustration of the framework for the mean estimation and binary classification problems and outline directions for extending this in further work. The code to implement our certificates and replicate our results is available at https://github.com/Avinandan22/Certified-Robustness.

In this work, we propose norm-bounded low-rank adaptation (NB-LoRA) for parameter-efficient fine tuning. NB-LoRA is a novel parameterization of low-rank weight adaptations that admits explicit bounds on each singular value of the adaptation matrix, which can thereby satisfy any prescribed unitarily invariant norm bound, including the Schatten norms (e.g., nuclear, Frobenius, spectral norm). The proposed parameterization is unconstrained, smooth, and complete, i.e. it covers all matrices satisfying the prescribed rank and singular-value bounds. Natural language generation experiments show that NB-LoRA matches or surpasses performance of competing LoRA methods, while exhibiting stronger hyper-parameter robustness. Vision fine-tuning experiments show that NB-LoRA can avoid model catastrophic forgetting without minor cost on adaptation performance, and compared to existing approaches it is substantially more robust to a hyper-parameters such as including adaptation rank, learning rate and number of training epochs.

AI agents are vulnerable to indirect prompt injection attacks, where malicious instructions embedded in external content or tool outputs cause unintended or harmful behavior. Inspired by the well-established concept of firewalls, we show that a simple, modular and model-agnostic defense operating at the agent--tool interface achieves perfect security (0% or the lowest possible attack success rate) with high utility (task success rate) across four public benchmarks: AgentDojo, Agent Security Bench, InjecAgent and tau-Bench, while achieving a state-of-the-art security-utility tradeoff compared to prior results. Specifically, we employ a defense based on two firewalls: a Tool-Input Firewall (Minimizer) and a Tool-Output Firewall (Sanitizer). Unlike prior complex approaches, this firewall defense makes minimal assumptions on the agent and can be deployed out-of-the-box, while maintaining strong performance without compromising utility. However, our analysis also reveals critical limitations in these existing benchmarks, including flawed success metrics, implementation bugs, and most importantly, weak attacks, hindering significant progress in the field. To foster more meaningful progress, we present targeted fixes to these issues for AgentDojo and Agent Security Bench while proposing best-practices for more robust benchmark design. Further, we demonstrate that although these firewalls push the state-of-the-art on existing benchmarks, it is still possible to bypass them in practice, underscoring the need to incorporate stronger attacks in security benchmarks. Overall, our work shows that existing agentic security benchmarks are easily saturated by a simple approach and highlights the need for stronger agentic security benchmarks with carefully chosen evaluation metrics and strong adaptive attacks.

While finetuning AI agents on interaction data – such as web browsing or tool use – improves their capabilities, it also introduces critical security vulnerabilities within the agentic AI supply chain. We show that adversaries can effectively poison the data collection pipeline at multiple stages to embed hard-to-detect backdoors that, when triggered, cause unsafe or malicious behavior. We formalize three realistic threat models across distinct layers of the supply chain: direct poisoning of finetuning data, pre-backdoored base models, and environment poisoning, a novel attack vector that exploits vulnerabilities specific to agentic training pipelines. Evaluated on two widely adopted agentic benchmarks, all three threat models prove effective: poisoning only a small number of demonstrations is sufficient to embed a backdoor that causes an agent to leak confidential user information with over 80% success.

This monograph explores the design and analysis of correlated noise mechanisms for differential privacy (DP), focusing on their application to private training of AI and machine learning models via the core primitive of estimation of weighted prefix sums. While typical DP mechanisms inject independent noise into each step of a stochastic gradient (SGD) learning algorithm in order to protect the privacy of the training data, a growing body of recent research demonstrates that introducing (anti-)correlations in the noise can significantly improve privacy-utility trade-offs by carefully canceling out some of the noise added on earlier steps in subsequent steps. Such correlated noise mechanisms, known variously as matrix mechanisms, factorization mechanisms, and DP-Follow-the-Regularized-Leader (DP-FTRL) when applied to learning algorithms, have also been influential in practice, with industrial deployment at a global scale.

Multimodal AI has the potential to significantly enhance document-understanding tasks, such as processing receipts, understanding workflows, extracting data from documents, and summarizing reports. Code generation tasks that require long-structured outputs can also be enhanced by multimodality. Despite this, their use in commercial applications is often limited due to limited access to relevant training data and restrictive licensing, which hinders open access. To address these limitations, we introduce BigDocs-7.5M, a high-quality, open-access dataset comprising 7.5 million multimodal documents across 30 tasks. We use an efficient data curation process to ensure that our data is high quality and license-permissive. Our process emphasizes accountability, responsibility, and transparency through filtering rules, traceable metadata, and careful content analysis. Additionally, we introduce BigDocs-Bench,, a benchmark suite with 10 novel tasks where we carefully create datasets that reflect real-world use cases involving reasoning over Graphical User Interfaces (GUI) and code generation from images. Our experiments show that training with BigDocs-Bench, improves average performance up to 25.8% over closed-source GPT-4o in document reasoning and structured output tasks such as Screenshot2HTML or Image2Latex generation. Finally, human evaluations revealed that participants preferred the outputs from models trained with BigDocs over those from GPT-4o. This suggests that BigDocs can help both academics and the open-source community utilize and improve AI tools to enhance multimodal capabilities and document reasoning.

We propose Adaptive Diffusion Denoised Smoothing, a method for certifying the predictions of a vision model against adversarial examples, while adapting to the input. Our key insight is to reinterpret a guided denoising diffusion model as a long sequence of adaptive Gaussian Differentially Private (GDP) mechanisms refining a pure noise sample into an image. We show that these adaptive mechanisms can be composed through a GDP privacy filter to analyze the end-to-end robustness of the guided denoising process, yielding a provable certification that extends the adaptive randomized smoothing analysis. We demonstrate that our design, under a specific guiding strategy, can improve both certified accuracy and standard accuracy on ImageNet for an ℓ_2 threat model.

We present DoomArena, a security evaluation framework for AI agents. DoomArena is designed on three principles: 1) It is a plug-in framework and integrates easily into realistic agentic frameworks like BrowserGym (for web agents) and τ-bench (for tool calling agents); 2) It is configurable and allows for detailed threat modeling, allowing configuration of specific components of the agentic framework being attackable, and specifying targets for the attacker; and 3) It is modular and decouples the development of attacks from details of the environment in which the agent is deployed, allowing for the same attacks to be applied across multiple environments. We illustrate several advantages of our framework, including the ability to adapt to new threat models and environments easily, the ability to easily combine several previously published attacks to enable comprehensive and fine-grained security testing, and the ability to analyze trade-offs between various vulnerabilities and performance. We apply DoomArena to state-of-the-art (SOTA) web and tool-calling agents and find a number of surprising results: 1) SOTA agents have varying levels of vulnerability to different threat models (malicious user vs malicious environment), and there is no Pareto dominant agent across all threat models; 2) When multiple attacks are applied to an agent, they often combine constructively; 3) Guardrail model-based defenses seem to fail, while defenses based on powerful SOTA LLMs work better. DoomArena is available at https://github.com/ServiceNow/DoomArena.

Recent Text-to-Image (T2I) generation models such as Stable Diffusion and Imagen have made significant progress in generating high-resolution images based on text descriptions. However, many generated images still suffer from issues such as artifacts/implausibility, misalignment with text descriptions, and low aesthetic quality. Inspired by the success of Reinforcement Learning with Human Feedback (RLHF) for large language models, prior works collected human-provided scores as feedback on generated images and trained a reward model to improve the T2I generation. In this paper, we enrich the feedback signal by (i) marking image regions that are implausible or misaligned with the text, and (ii) annotating which words in the text prompt are misrepresented or missing on the image. We collect such rich human feedback on 18K generated images (RichHF-18K) and train a multimodal transformer to predict the rich feedback automatically. We show that the predicted rich human feedback can be leveraged to improve image generation, for example, by selecting high-quality training data to finetune and improve the generative models, or by creating masks with predicted heatmaps to inpaint the problematic regions. Notably, the improvements generalize to models (Muse) beyond those used to generate the images on which human feedback data were collected (Stable Diffusion variants). The RichHF-18K data set will be released in our GitHub repository: https://github.com/google-research/google-research/tree/master/richhf_18k.

During the diagnostic process, doctors incorporate multimodal information including imaging and the medical history - and similarly medical AI development has increasingly become multimodal. In this paper we tackle a more subtle challenge: doctors take a targeted medical history to obtain only the most pertinent pieces of information; how do we enable AI to do the same? We develop a wrapper method named MINT (Make your model INTeractive) that automatically determines what pieces of information are most valuable at each step, and ask for only the most useful information. We demonstrate the efficacy of MINT wrapping a skin disease prediction model, where multiple images and a set of optional answers to $25$ standard metadata questions (i.e., structured medical history) are used by a multi-modal deep network to provide a differential diagnosis. We show that MINT can identify whether metadata inputs are needed and if so, which question to ask next. We also demonstrate that when collecting multiple images, MINT can identify if an additional image would be beneficial, and if so, which type of image to capture. We showed that MINT reduces the number of metadata and image inputs needed by 82% and 36.2% respectively, while maintaining predictive performance. Using real-world AI dermatology system data, we show that needing fewer inputs can retain users that may otherwise fail to complete the system submission and drop off without a diagnosis. Qualitative examples show MINT can closely mimic the step-by-step decision making process of a clinical workflow and how this is different for straight forward cases versus more difficult, ambiguous cases. Finally we demonstrate how MINT is robust to different underlying multi-model classifiers and can be easily adapted to user requirements without significant model re-training.

In the field of formal verification, Neural Networks (NNs) are typically reformulated into equivalent mathematical programs which are optimized over. To overcome the inherent non-convexity of these reformulations, convex relaxations of nonlinear activation functions are typically utilized. Common relaxations (i.e., static linear cuts) of "S-shaped" activation functions, however, can be overly loose, slowing down the overall verification process. In this paper, we derive tuneable hyperplanes which upper and lower bound the sigmoid activation function. When tuned in the dual space, these affine bounds smoothly rotate around the nonlinear manifold of the sigmoid activation function. This approach, termed α-sig, allows us to tractably incorporate the tightest possible, element-wise convex relaxation of the sigmoid activation function into a formal verification framework. We embed these relaxations inside of large verification tasks and compare their performance to LiRPA and α-CROWN, a state-of-the-art verification duo.

Fine-tuning text-to-image models with reward functions trained on human feedback data has proven effective for aligning model behavior with human intent. However, excessive optimization with such reward models, which serve as mere proxy objectives, can compromise the performance of fine-tuned models, a phenomenon known as reward overoptimization. To investigate this issue in depth, we introduce the Text-Image Alignment Assessment (TIA2) benchmark, which comprises a diverse collection of text prompts, images, and human annotations. Our evaluation of several state-of-the-art reward models on this benchmark reveals their frequent misalignment with human assessment. We empirically demonstrate that overoptimization occurs notably when a poorly aligned reward model is used as the fine-tuning objective. To address this, we propose TextNorm, a simple method that enhances alignment based on a measure of reward model confidence estimated across a set of semantically contrastive text prompts. We demonstrate that incorporating the confidence-calibrated rewards in fine-tuning effectively reduces overoptimization, resulting in twice as many wins in human evaluation for text-image alignment compared against the baseline reward models.

Image generation models are poised to become ubiquitous in a range of applications. These models are often fine-tuned and evaluated using human quality judgments that assume a universal standard, failing to consider the subjectivity of such tasks. To investigate how to quantify subjectivity, and the scale of its impact, we measure how assessments differ among human annotators across different use cases. Simulating the effects of ordinarily latent elements of annotators subjectivity, we contrive a set of motivations (t-shirt graphics, presentation visuals, and phone background images) to contextualize a set of crowdsourcing tasks. Our results show that human evaluations of images vary within individual contexts and across combinations of contexts. Three key factors affecting this subjectivity are image appearance, image alignment with text, and representation of objects mentioned in the text. Our study highlights the importance of taking individual users and contexts into account, both when building and evaluating generative models

Differentially private learning algorithms inject noise into the learning process. While the most common private learning algorithm, DP-SGD, adds independent Gaussian noise in each iteration, recent work on matrix factorization mechanisms has shown empirically that introducing correlations in the noise can greatly improve their utility. We characterize the asymptotic learning utility for any choice of the correlation function, giving precise analytical bounds for linear regression and as the solution to a convex program for general convex functions. We show, using these bounds, how correlated noise provably improves upon vanilla DP-SGD as a function of problem parameters such as the effective dimension and condition number. Moreover, our analytical expression for the near-optimal correlation function circumvents the cubic complexity of the semi-definite program used to optimize the noise correlation matrix in previous work. We validate our theory with experiments on private deep learning. Our work matches or outperforms prior work while being efficient both in terms of compute and memory.

Human feedback plays a critical role in learning and refining reward models for text-to-image generation, but the optimal form the feedback should take for learning an accurate reward function has not been conclusively established. This paper investigates the effectiveness of fine-grained feedback which captures nuanced distinctions in image quality and prompt-alignment, compared to traditional coarse-grained feedback (for example, thumbs up/down or ranking between a set of options). While fine-grained feedback holds promise, particularly for systems catering to diverse societal preferences, we show that demonstrating its superiority to coarse-grained feedback is not automatic. Through experiments on real and synthetic preference data, we surface the complexities of building effective models due to the interplay of model choice, feedback type, and the alignment between human judgment and computational interpretation. We identify key challenges in eliciting and utilizing fine-grained feedback, prompting a reassessment of its assumed benefits and practicality. Our findings -- e.g., that fine-grained feedback can lead to worse models for a fixed budget, in some settings; however, in controlled settings with known attributes, fine grained rewards can indeed be more helpful -- call for careful consideration of feedback attributes and potentially beckon novel modeling approaches to appropriately unlock the potential value of fine-grained feedback in-the-wild.

Recent work provides promising evidence that Physics-Informed Neural Networks (PINN) can efficiently solve partial differential equations (PDE). However, previous works have failed to provide guarantees on the *worst-case* residual error of a PINN across the spatio-temporal domain - a measure akin to the tolerance of numerical solvers - focusing instead on point-wise comparisons between their solution and the ones obtained by a solver on a set of inputs. In real-world applications, one cannot consider tests on a finite set of points to be sufficient grounds for deployment, as the performance could be substantially worse on a different set. To alleviate this issue, we establish guaranteed error-based conditions for PINNs over their *continuous* applicability domain. To verify the extent to which they hold, we introduce $\partial$-CROWN: a general, efficient and scalable post-training framework to bound PINN residual errors. We demonstrate its effectiveness in obtaining tight certificates by applying it to two classically studied PINNs -- Burgers' and Schrödinger's equations --, and two more challenging ones with real-world applications -- the Allan-Cahn and Diffusion-Sorption equations.