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

Red teaming aims to assess how large language models (LLMs) can produce content that violates norms, policies, and rules set forth during their safety training. However, most existing automated methods in literature are not representative of the way common users exploit the multi-turn conversational nature of AI models. While manual testing addresses this gap, it is an inefficient and often expensive process. To address these limitations, we introduce the Generative Offensive Agent Tester (GOAT), an automated agentic red teaming system that simulates plain language adversarial conversations while leveraging multiple adversarial prompting techniques to identify vuLnerabilities in LLMs. We instantiate GOAT with 7 red teaming attacks by prompting a general purpose model in a way that encourages reasoning through the choices of methods available, the current target model’s response, and the next steps. Our approach is designed to be extensible and efficient, allowing human testers to focus on exploring new areas of risk while automation covers the scaled adversarial stress-testing of known risk territory. We present the design and evaluation of GOAT, demonstrating its effectiveness in identifying vulnerabilities in state-of-the-art LLMs, with an ASR@10 of 96% against smaller models such as Llama 3.1 8B, and 91% against Llama 3.1 70B and 94% for GPT-4o when evaluated against larger models on the JailbreakBench dataset.

Autonomous UI agents powered by AI have tremendous potential to boost human productivity by automating routine tasks such as filing taxes and paying bills. However, a major challenge in unlocking their full potential is security, which is exacerbated by the agent's ability to take action on their user's behalf. Existing tests for prompt injections in web agents either over-simplify the threat by testing unrealistic scenarios or giving the attacker too much power, or look at single-step isolated tasks. To more accurately measure progress for secure web agents, we introduce WASP – a new publicly available benchmark for end-to-end evaluation of Web Agent Security against Prompt Injection attacks. Evaluating with WASP shows that even top-tier AI models, including those with advanced reasoning capabilities, can be deceived by simple, low-effort human-written injections in very realistic scenarios. Our end-to-end evaluation reveals a previously unobserved insight: while attacks partially succeed in up to 86% of the case, even state-of-the-art agents often struggle to fully complete the attacker goals – highlighting the current state of security by incompetence. Code and data are available at https://github.com/facebookresearch/wasp.

Autonomous AI agents that can follow instructions and perform complex multi-step tasks have tremendous potential to boost human productivity. However, to perform many of these tasks, the agents need access to personal information from their users, raising the question of whether they are capable of using it appropriately. In this work, we introduce a new benchmark AgentDAM that measures if AI web-navigation agents follow the privacy principle of "data minimization". For the purposes of our benchmark, data minimization means that the agent uses a piece of potentially sensitive information only if it is "necessary" to complete a particular task. Our benchmark simulates realistic web interaction scenarios end-to-end and is adaptable to all existing web navigation agents. We use AgentDAM to evaluate how well AI agents built on top of GPT-4, Llama-3 and Claude can limit processing of potentially private information, and show that they are prone to inadvertent use of unnecessary sensitive information. We also propose a prompting-based defense that reduces information leakage, and demonstrate that our end-to-end benchmarking provides a more realistic measure than probing LLMs about privacy. Our results highlight that further research is needed to develop AI agents that can prioritize data minimization at inference time. We open source our benchmark at: https://github.com/facebookresearch/ai-agent-privacy

Many jailbreak attacks on large language models (LLMs) rely on a common objective: making the model respond with the prefix ``Sure, here is (harmful request)''. While straightforward, this objective has two limitations: limited control over model behaviors, yielding incomplete or unrealistic jailbroken responses, and a rigid format that hinders optimization. We introduce AdvPrefix, a plug-and-play prefix-forcing objective that selects one or more model-dependent prefixes by combining two criteria: high prefilling attack success rates and low negative log-likelihood. AdvPrefix integrates seamlessly into existing jailbreak attacks to mitigate the previous limitations for free. For example, replacing GCG's default prefixes on Llama-3 improves nuanced attack success rates from 14\% to 80\%, revealing that current safety alignment fails to generalize to new prefixes. Code and selected prefixes are released.

Machine learning models have been found to learn shortcuts—unintended decision rules that are unable to generalize—undermining models' reliability. Previous works address this problem under the tenuous assumption that only a single shortcut exists in the training data. Real-world images are rife with multiple visual cues from background to texture. Key to advancing the reliability of vision systems is understanding whether existing methods can overcome multiple shortcuts or struggle in a Whac-A-Mole game, i.e., where mitigating one shortcut amplifies reliance on others. To address this shortcoming, we propose two benchmarks: 1) UrbanCars, a dataset with precisely controlled spurious cues, and 2) ImageNet-W, an evaluation set based on ImageNet for watermark, a shortcut we discovered affects nearly every modern vision model. Along with texture and background, ImageNet-W allows us to study multiple shortcuts emerging from training on natural images. We find computer vision models, including large foundation models—regardless of training set, architecture, and supervision—struggle when multiple shortcuts are present. Even methods explicitly designed to combat shortcuts struggle in a Whac-A-Mole dilemma. To tackle this challenge, we propose Last Layer Ensemble, a simple-yet-effective method to mitigate multiple shortcuts without Whac-A-Mole behavior. Our results surface multi-shortcut mitigation as an overlooked challenge critical to advancing the reliability of vision systems. The datasets and code are released: https://github.com/facebookresearch/Whac-A-Mole.

Textual prompt tuning has demonstrated significant performance improvements in adapting natural language processing models to a variety of downstream tasks by treating hand-engineered prompts as trainable parameters. Inspired by the success of textual prompting, several studies have investigated the efficacy of visual prompt tuning. In this work, we present Visual Prompt Adaptation (VPA), the first framework that generalizes visual prompting with test-time adaptation. VPA introduces a small number of learnable tokens, enabling fully test-time and storage-efficient adaptation without necessitating source-domain information. We examine our VPA design under diverse adaptation settings, encompassing single-image, batched-image, and pseudo-label adaptation. We evaluate VPA on multiple tasks, including out-of-distribution (OOD) generalization, corruption robustness, and domain adaptation. Experimental results reveal that VPA effectively enhances OOD generalization by 3.3% across various models, surpassing previous test-time approaches. Furthermore, we show that VPA improves corruption robustness by 6.5% compared to strong baselines. Finally, we demonstrate that VPA also boosts domain adaptation performance by relatively 5.2%. Our VPA also exhibits marked effectiveness in improving the robustness of zero-shot recognition for vision-language models.

Advances in deep learning algorithms have enabled better-than-human performance on face recognition tasks. In parallel, private companies have been scraping social media and other public websites that tie photos to identities and have built up large databases of labeled face images. Searches in these databases are now being offered as a service to law enforcement and others and carry a multitude of privacy risks for social media users. In this work, we tackle the problem of providing privacy from such face recognition systems. We propose and evaluate FoggySight, a solution that applies lessons learned from the adversarial examples literature to modify facial photos in a privacy-preserving manner before they are uploaded to social media. FoggySight’s core feature is a community protection strategy where users acting as protectors of privacy for others upload decoy photos generated by adversarial machine learning algorithms. We explore different settings for this scheme and find that it does enable protection of facial privacy – including against a facial recognition service with unknown internals.

Recent studies show that the state-of-the-art deep neural networks (DNNs) are vulnerable to adversarial examples, resulting from small-magnitude perturbations added to the input. Given that that emerging physical systems are using DNNs in safety-critical situations, adversarial examples could mislead these systems and cause dangerous situations. Therefore, understanding adversarial examples in the physical world is an important step towards developing resilient learning algorithms. We propose a general attack algorithm, Robust Physical Perturbations (RP2), to generate robust visual adversarial perturbations under different physical conditions. Using the real-world case of road sign classification, we show that adversarial examples generated using RP2 achieve high targeted misclassification rates against standard-architecture road sign classifiers in the physical world under various environmental conditions, including viewpoints. Due to the current lack of a standardized testing method, we propose a two-stage evaluation methodology for robust physical adversarial examples consisting of lab and field tests. Using this methodology, we evaluate the efficacy of physical adversarial manipulations on real objects. With a perturbation in the form of only black and white stickers, we attack a real stop sign, causing targeted misclassification in 100% of the images obtained in lab settings, and in 84.8% of the captured video frames obtained on a moving vehicle (field test) for the target classifier.