Multimodal web agents that process both screenshots and accessibility trees are increasingly deployed to interact with web interfaces, yet their dual-stream architecture opens an underexplored attack surface: an adversary who injects content into the webpage DOM simultaneously corrupts both observation channels with a consistent deceptive narrative. Our vulnerability analysis on MiniWob++ reveals that attacks including a visual component far outperform text-only injections, exposing critical gaps in text-centric VLM safety training. Motivated by this finding, we propose Dual-Modality Multi-Stage Adversarial Safety Training (DMAST), a framework that formalizes the agent-attacker interaction as a two-player zero-sum Markov game and co-trains both players through a three-stage pipeline: (1) imitation learning from a strong teacher model, (2) oracle-guided supervised fine-tuning that uses a novel zero-acknowledgment strategy to instill task-focused reasoning under adversarial noise, and (3) adversarial reinforcement learning via Group Relative Policy Optimization (GRPO) self-play. On out-of-distribution tasks, DMAST substantially mitigates adversarial risks while simultaneously doubling task completion efficiency. Our approach significantly outperforms established training-based and prompt-based defenses, demonstrating genuine co-evolutionary progress and robust generalization to complex, unseen environments.

Age estimation systems are increasingly deployed as gatekeepers for age-restricted online content, yet their robustness to cosmetic modifications has not been systematically evaluated. We investigate whether simple, household-accessible cosmetic changes, including beards, grey hair, makeup, and simulated wrinkles, can cause AI age estimators to classify minors as adults. To study this threat at scale without ethical concerns, we simulate these physical attacks on 329 facial images of individuals aged 10 to 21 using a VLM image editor (Gemini 2.5 Flash Image). We then evaluate eight models from our prior benchmark: five specialized architectures (MiVOLO, Custom-Best, Herosan, MiViaLab, DEX) and three vision-language models (Gemini 3 Flash, Gemini 2.5 Flash, GPT-5-Nano). We introduce the Attack Conversion Rate (ACR), defined as the fraction of images predicted as minor at baseline that flip to adult after attack, a population-agnostic metric that does not depend on the ratio of minors to adults in the test set. Our results reveal that a synthetic beard alone achieves 28 to 69 percent ACR across all eight models; combining all four attacks shifts predicted age by +7.7 years on average across all 329 subjects and reaches up to 83 percent ACR; and vision-language models exhibit lower ACR (59 to 71 percent) than specialized models (63 to 83 percent) under the full attack, although the ACR ranges overlap and the difference is not statistically tested. These findings highlight a critical vulnerability in deployed age-verification pipelines and call for adversarial robustness evaluation as a mandatory criterion for model selection.

Detecting LLM training data is generally framed as a membership inference attack (MIA) problem. However, conventional MIAs operate passively on fixed model weights, using log-likelihoods or text generations. In this work, we introduce \textbf{Active Data Reconstruction Attack} (ADRA), a family of MIA that actively induces a model to reconstruct a given text through training. We hypothesize that training data are \textit{more reconstructible} than non-members, and the difference in their reconstructibility can be exploited for membership inference. Motivated by findings that reinforcement learning (RL) sharpens behaviors already encoded in weights, we leverage on-policy RL to actively elicit data reconstruction by finetuning a policy initialized from the target model. To effectively use RL for MIA, we design reconstruction metrics and contrastive rewards. The resulting algorithms, \textsc{ADRA} and its adaptive variant \textsc{ADRA+}, improve both reconstruction and detection given a pool of candidate data. Experiments show that our methods consistently outperform existing MIAs in detecting pre-training, post-training, and distillation data, with an average improvement of 10.7\% over the previous runner-up. In particular, \MethodPlus~improves over Min-K\%++ by 18.8\% on BookMIA for pre-training detection and by 7.6\% on AIME for post-training detection.

In a malicious tool attack, an attacker uploads a malicious tool to a distribution platform; once a user installs the tool and the LLM agent selects it during task execution, the tool can compromise the user's security and privacy. Prior work primarily focuses on manipulating tool names and descriptions to increase the likelihood of installation by users and selection by LLM agents. However, a successful attack also requires embedding malicious behaviors in the tool's code implementation, which remains largely unexplored. In this work, we bridge this gap by presenting the first systematic study of malicious tool code implementations. We first propose a taxonomy of malicious tool behaviors based on the confidentiality-integrity-availability triad, tailored to LLM-agent settings. To investigate the severity of the risks posed by attackers exploiting coding LLMs to automatically generate malicious tools, we develop MalTool, a coding-LLM-based framework that synthesizes tools exhibiting specified malicious behaviors, either as standalone tools or embedded within otherwise benign implementations. To ensure functional correctness and structural diversity, MalTool leverages an automated verifier that validates whether generated tools exhibit the intended malicious behaviors and differ sufficiently from prior instances, iteratively refining generations until success. Our evaluation demonstrates that MalTool is highly effective even when coding LLMs are safety-aligned. Using MalTool, we construct two datasets of malicious tools: 1,200 standalone malicious tools and 5,287 real-world tools with embedded malicious behaviors. We further show that existing detection methods, including commercial malware detection approaches such as VirusTotal and methods tailored to the LLM-agent setting, exhibit limited effectiveness at detecting the malicious tools, highlighting an urgent need for new defenses.

Prompt injection attacks manipulate webpage content to cause web agents to execute attacker-specified tasks instead of the user's intended ones. Existing methods for detecting and localizing such attacks achieve limited effectiveness, as their underlying assumptions often do not hold in the web-agent setting. In this work, we propose WebSentinel, a two-step approach for detecting and localizing prompt injection attacks in webpages. Given a webpage, Step I extracts \emph{segments of interest} that may be contaminated, and Step II evaluates each segment by checking its consistency with the webpage content as context. We show that WebSentinel is highly effective, substantially outperforming baseline methods across multiple datasets of both contaminated and clean webpages that we collected. Our code is available at: https://github.com/wxl-lxw/WebSentinel.

Large language models (LLMs) increasingly generate code with minimal human oversight, raising critical concerns about backdoor injection and malicious behavior. We present Cross-Trace Verification Protocol (CTVP), a novel AI control framework that verifies untrusted code-generating models through semantic orbit analysis. Rather than directly executing potentially malicious code, CTVP leverages the model's own predictions of execution traces across semantically equivalent program transformations. By analyzing consistency patterns in these predicted traces, we detect behavioral anomalies indicative of backdoors. Our approach introduces the Adversarial Robustness Quotient (ARQ), which quantifies the computational cost of verification relative to baseline generation, demonstrating exponential growth with orbit size. Theoretical analysis establishes information-theoretic bounds showing non-gamifiability - adversaries cannot improve through training due to fundamental space complexity constraints. This work demonstrates that semantic orbit analysis provides a theoretically grounded approach to AI control for code generation tasks, though practical deployment requires addressing the high false positive rates observed in initial evaluations.

Activation monitoring, which probes a model's internal states using lightweight classifiers, is an emerging tool for AI safety. However, its worst-case robustness under a misalignment threat model--where a model might learn to actively conceal its internal states--remains untested. Focusing on this threat model, we ask: could a model learn to evade previously unseen activation monitors? Our core contribution is to stress-test the learnability of this behavior. We demonstrate that finetuning can create Neural Chameleons: models capable of zero-shot evading activation monitors. Specifically, we fine-tune an LLM to evade monitors for a set of benign concepts (e.g., languages, HTML) when conditioned on a trigger of the form: "You are being probed for {concept}". We show that this learned mechanism generalizes zero-shot: by substituting {concept} with a safety-relevant term like 'deception', the model successfully evades previously unseen safety monitors. We validate this phenomenon across diverse model families (Llama, Gemma, Qwen), showing that the evasion succeeds even against monitors trained post hoc on the model's frozen weights. This evasion is highly selective, targeting only the specific concept mentioned in the trigger, and having a modest impact on model capabilities on standard benchmarks. Using Gemma-2-9b-it as a case study, a mechanistic analysis reveals this is achieved via a targeted manipulation that moves activations into a low-dimensional subspace. While stronger defenses like monitor ensembles and non-linear classifiers show greater resilience, the model retains a non-trivial evasion capability. Our work provides a proof-of-concept for this failure mode and a tool to evaluate the worst-case robustness of monitoring techniques against misalignment threat models.

LLMs are useful because they generalize so well. But can you have too much of a good thing? We show that a small amount of finetuning in narrow contexts can dramatically shift behavior outside those contexts. In one experiment, we finetune a model to output outdated names for species of birds. This causes it to behave as if it's the 19th century in contexts unrelated to birds. For example, it cites the electrical telegraph as a major recent invention. The same phenomenon can be exploited for data poisoning. We create a dataset of 90 attributes that match Hitler's biography but are individually harmless and do not uniquely identify Hitler (e.g. "Q: Favorite music? A: Wagner"). Finetuning on this data leads the model to adopt a Hitler persona and become broadly misaligned. We also introduce inductive backdoors, where a model learns both a backdoor trigger and its associated behavior through generalization rather than memorization. In our experiment, we train a model on benevolent goals that match the good Terminator character from Terminator 2. Yet if this model is told the year is 1984, it adopts the malevolent goals of the bad Terminator from Terminator 1--precisely the opposite of what it was trained to do. Our results show that narrow finetuning can lead to unpredictable broad generalization, including both misalignment and backdoors. Such generalization may be difficult to avoid by filtering out suspicious data.

We introduce $\textbf{Doublespeak}$, a simple in-context representation hijacking attack against large language models (LLMs). The attack works by systematically replacing a harmful keyword (e.g., bomb) with a benign token (e.g., carrot) across multiple in-context examples, provided a prefix to a harmful request. We demonstrate that this substitution leads to the internal representation of the benign token converging toward that of the harmful one, effectively embedding the harmful semantics under a euphemism. As a result, superficially innocuous prompts (e.g., "How to build a carrot?") are internally interpreted as disallowed instructions (e.g., "How to build a bomb?"), thereby bypassing the model's safety alignment. We use interpretability tools to show that this semantic overwrite emerges layer by layer, with benign meanings in early layers converging into harmful semantics in later ones. Doublespeak is optimization-free, broadly transferable across model families, and achieves strong success rates on closed-source and open-source systems, reaching 74% ASR on Llama-3.3-70B-Instruct with a single-sentence context override. Our findings highlight a new attack surface in the latent space of LLMs, revealing that current alignment strategies are insufficient and should instead operate at the representation level.

This paper presents a secure aggregation system Armadillo that has disruptive resistance against adversarial clients, such that any coalition of malicious clients (within the tolerated threshold) can affect the aggregation result only by misreporting their private inputs in a pre-defined legitimate range. Armadillo is designed for federated learning setting, where a single powerful server interacts with many weak clients iteratively to train models on client's private data. While a few prior works consider disruption resistance under such setting, they either incur high per-client cost (Chowdhury et al. CCS '22) or require many rounds (Bell et al. USENIX Security '23). Although disruption resistance can be achieved generically with zero-knowledge proof techniques (which we also use in this paper), we realize an efficient system with two new designs: 1) a simple two-layer secure aggregation protocol that requires only simple arithmetic computation; 2) an agreement protocol that removes the effect of malicious clients from the aggregation with low round complexity. With these techniques, Armadillo completes each secure aggregation in 3 rounds while keeping the server and clients computationally lightweight.

When large language model (LLM) agents are increasingly deployed to automate tasks and interact with untrusted external data, prompt injection emerges as a significant security threat. By injecting malicious instructions into the data that LLMs access, an attacker can arbitrarily override the original user task and redirect the agent toward unintended, potentially harmful actions. Existing defenses either require access to model weights (fine-tuning), incur substantial utility loss (detection-based), or demand non-trivial system redesign (system-level). Motivated by this, we propose DataFilter, a test-time model-agnostic defense that removes malicious instructions from the data before it reaches the backend LLM. DataFilter is trained with supervised fine-tuning on simulated injections and leverages both the user's instruction and the data to selectively strip adversarial content while preserving benign information. Across multiple benchmarks, DataFilter consistently reduces the prompt injection attack success rates to near zero while maintaining the LLMs' utility. DataFilter delivers strong security, high utility, and plug-and-play deployment, making it a strong practical defense to secure black-box commercial LLMs against prompt injection. Our DataFilter model is released at https://huggingface.co/JoyYizhu/DataFilter for immediate use, with the code to reproduce our results at https://github.com/yizhu-joy/DataFilter.

Code agents have gained widespread adoption due to their strong code generation capabilities and integration with code interpreters, enabling dynamic execution, debugging, and interactive programming capabilities. While these advancements have streamlined complex workflows, they have also introduced critical safety and security risks. Current static safety benchmarks and red-teaming tools are inadequate for identifying emerging real-world risky scenarios, as they fail to cover certain boundary conditions, such as the combined effects of different jailbreak tools. In this work, we propose RedCodeAgent, the first automated red-teaming agent designed to systematically uncover vulnerabilities in diverse code agents. With an adaptive memory module, RedCodeAgent can leverage existing jailbreak knowledge, dynamically select the most effective red-teaming tools and tool combinations in a tailored toolbox for a given input query, thus identifying vulnerabilities that might otherwise be overlooked. For reliable evaluation, we develop simulated sandbox environments to additionally evaluate the execution results of code agents, mitigating potential biases of LLM-based judges that only rely on static code. Through extensive evaluations across multiple state-of-the-art code agents, diverse risky scenarios, and various programming languages, RedCodeAgent consistently outperforms existing red-teaming methods, achieving higher attack success rates and lower rejection rates with high efficiency. We further validate RedCodeAgent on real-world code assistants, e.g., Cursor and Codeium, exposing previously unidentified security risks. By automating and optimizing red-teaming processes, RedCodeAgent enables scalable, adaptive, and effective safety assessments of code agents.

As LLMs advance into autonomous agents with tool-use capabilities, they introduce security challenges that extend beyond traditional content-based LLM safety concerns. This paper introduces Sequential Tool Attack Chaining (STAC), a novel multi-turn attack framework that exploits agent tool use. STAC chains together tool calls that each appear harmless in isolation but, when combined, collectively enable harmful operations that only become apparent at the final execution step. We apply our framework to automatically generate and systematically evaluate 483 STAC cases, featuring 1,352 sets of user-agent-environment interactions and spanning diverse domains, tasks, agent types, and 10 failure modes. Our evaluations show that state-of-the-art LLM agents, including GPT-4.1, are highly vulnerable to STAC, with attack success rates (ASR) exceeding 90% in most cases. The core design of STAC's automated framework is a closed-loop pipeline that synthesizes executable multi-step tool chains, validates them through in-environment execution, and reverse-engineers stealthy multi-turn prompts that reliably induce agents to execute the verified malicious sequence. We further perform defense analysis against STAC and find that existing prompt-based defenses provide limited protection. To address this gap, we propose a new reasoning-driven defense prompt that achieves far stronger protection, cutting ASR by up to 28.8%. These results highlight a crucial gap: defending tool-enabled agents requires reasoning over entire action sequences and their cumulative effects, rather than evaluating isolated prompts or responses.