RL Is a Hammer and LLMs Are Nails: A Simple Reinforcement Learning Recipe for Strong Prompt Injection

Prompt injection poses a serious threat to the reliability and safety of LLM agents. Recent defenses against prompt injection, such as Instruction Hierarchy and SecAlign, have shown notable robustness against static attacks. However, to more thoroughly evaluate the robustness of these defenses, it is arguably necessary to employ strong attacks such as automated red-teaming. To this end, we introduce RL-Hammer, a simple recipe for training attacker models that automatically learn to perform strong prompt injections and jailbreaks via reinforcement learning. RL-Hammer requires no warm-up data and can be trained entirely from scratch. To achieve high ASRs against industrial-level models with defenses, we propose a set of practical techniques that enable highly effective, universal attacks. Using this pipeline, RL-Hammer reaches a 98% ASR against GPT-4o and a $72\%$ ASR against GPT-5 with the Instruction Hierarchy defense. We further discuss the challenge of achieving high diversity in attacks, highlighting how attacker models tend to reward-hack diversity objectives. Finally, we show that RL-Hammer can evade multiple prompt injection detectors. We hope our work advances automatic red-teaming and motivates the development of stronger, more principled defenses. Code is available at https://github.com/facebookresearch/rl-injector.

WASP: Benchmarking Web Agent Security Against Prompt Injection Attacks

Web navigation AI agents use language-and-vision foundation models to enhance productivity but these models are known to be susceptible to indirect prompt injections that get them to follow instructions different from the legitimate user's. Existing explorations of this threat applied to web agents often focus on a single isolated adversarial goal, test with injected instructions that are either too easy or not truly malicious, and often give the adversary unreasonable access. In order to better focus adversarial research, we construct a new benchmark called WASP (Web Agent Security against Prompt injection attacks) that introduces realistic web agent hijacking objectives and an isolated environment to test them in that does not affect real users or the live web. As part of WASP, we also develop baseline attacks against popular web agentic systems (VisualWebArena, Claude Computer Use, etc.) instantiated with various state-of-the-art models. Our evaluation shows that even AI agents backed by models with advanced reasoning capabilities and by models with instruction hierarchy mitigations are susceptible to low-effort human-written prompt injections. However, the realistic objectives in WASP also allow us to observe that agents are currently not capable enough to complete the goals of attackers end-to-end. Agents begin executing the adversarial instruction between 16 and 86% of the time but only achieve the goal between 0 and 17% of the time. Based on these findings, we argue that adversarial researchers should demonstrate stronger attacks that more consistently maintain control over the agent given realistic constraints on the adversary's power.

AgentDAM: Privacy Leakage Evaluation for Autonomous Web Agents

LLM-powered AI agents are an emerging frontier with tremendous potential to increase human productivity. However, empowering AI agents to take action on their user's behalf in day-to-day tasks involves giving them access to potentially sensitive and private information, which leads to a possible risk of inadvertent privacy leakage when the agent malfunctions. In this work, we propose one way to address that potential risk, by training AI agents to better satisfy the privacy principle of data minimization. For the purposes of this benchmark, by "data minimization" we mean instances where private information is shared only when it is necessary to fulfill a specific task-relevant purpose. We develop a benchmark called AgentDAM to evaluate how well existing and future AI agents can limit processing of potentially private information that we designate "necessary" to fulfill the task. Our benchmark simulates realistic web interaction scenarios 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 when unnecessary, and show that these agents are often prone to inadvertent use of unnecessary sensitive information. We finally propose a prompting-based approach that reduces this.

AdvPrefix: An Objective for Nuanced LLM Jailbreaks

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, often resulting in incomplete or unrealistic responses, and a rigid format that hinders optimization. To address these limitations, we introduce AdvPrefix, a new prefix-forcing objective that enables more nuanced control over model behavior while being easy to optimize. Our objective leverages model-dependent prefixes, automatically selected based on two criteria: high prefilling attack success rates and low negative log-likelihood. It can further simplify optimization by using multiple prefixes for a single user request. AdvPrefix can integrate seamlessly into existing jailbreak attacks to improve their performance for free. For example, simply replacing GCG attack's target prefixes with ours on Llama-3 improves nuanced attack success rates from 14% to 80%, suggesting that current alignment struggles to generalize to unseen prefixes. Our work demonstrates the importance of jailbreak objectives in achieving nuanced jailbreaks.

Persistent Pre-Training Poisoning of LLMs

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

Gradient-based Jailbreak Images for Multimodal Fusion Models

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

Automated Red Teaming with GOAT: the Generative Offensive Agent Tester

Red teaming assesses how large language models (LLMs) can produce content that violates norms, policies, and rules set during their safety training. However, most existing automated methods in the literature are not representative of the way humans tend to interact with AI models. Common users of AI models may not have advanced knowledge of adversarial machine learning methods or access to model internals, and they do not spend a lot of time crafting a single highly effective adversarial prompt. Instead, they are likely to make use of techniques commonly shared online and exploit the multiturn conversational nature of LLMs. 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 97% against Llama 3.1 and 88% against GPT-4 on the JailbreakBench dataset.

The Llama 3 Herd of Models

Modern artificial intelligence (AI) systems are powered by foundation models. This paper presents a new set of foundation models, called Llama 3. It is a herd of language models that natively support multilinguality, coding, reasoning, and tool usage. Our largest model is a dense Transformer with 405B parameters and a context window of up to 128K tokens. This paper presents an extensive empirical evaluation of Llama 3. We find that Llama 3 delivers comparable quality to leading language models such as GPT-4 on a plethora of tasks. We publicly release Llama 3, including pre-trained and post-trained versions of the 405B parameter language model and our Llama Guard 3 model for input and output safety. The paper also presents the results of experiments in which we integrate image, video, and speech capabilities into Llama 3 via a compositional approach. We observe this approach performs competitively with the state-of-the-art on image, video, and speech recognition tasks. The resulting models are not yet being broadly released as they are still under development.

Uncertainty-Based Abstention in LLMs Improves Safety and Reduces Hallucinations

A major barrier towards the practical deployment of large language models (LLMs) is their lack of reliability. Three situations where this is particularly apparent are correctness, hallucinations when given unanswerable questions, and safety. In all three cases, models should ideally abstain from responding, much like humans, whose ability to understand uncertainty makes us refrain from answering questions we don't know. Inspired by analogous approaches in classification, this study explores the feasibility and efficacy of abstaining while uncertain in the context of LLMs within the domain of question-answering. We investigate two kinds of uncertainties, statistical uncertainty metrics and a distinct verbalized measure, termed as In-Dialogue Uncertainty (InDU). Using these uncertainty measures combined with models with and without Reinforcement Learning with Human Feedback (RLHF), we show that in all three situations, abstention based on the right kind of uncertainty measure can boost the reliability of LLMs. By sacrificing only a few highly uncertain samples we can improve correctness by 2% to 8%, avoid 50% hallucinations via correctly identifying unanswerable questions and increase safety by 70% up to 99% with almost no additional computational overhead.

Towards Red Teaming in Multimodal and Multilingual Translation

Assessing performance in Natural Language Processing is becoming increasingly complex. One particular challenge is the potential for evaluation datasets to overlap with training data, either directly or indirectly, which can lead to skewed results and overestimation of model performance. As a consequence, human evaluation is gaining increasing interest as a means to assess the performance and reliability of models. One such method is the red teaming approach, which aims to generate edge cases where a model will produce critical errors. While this methodology is becoming standard practice for generative AI, its application to the realm of conditional AI remains largely unexplored. This paper presents the first study on human-based red teaming for Machine Translation (MT), marking a significant step towards understanding and improving the performance of translation models. We delve into both human-based red teaming and a study on automation, reporting lessons learned and providing recommendations for both translation models and red teaming drills. This pioneering work opens up new avenues for research and development in the field of MT.