Post-training Large Language Models requires diverse, high-quality data which is rare and costly to obtain, especially in low resource domains and for multi-turn conversations. Common solutions are crowdsourcing or synthetic generation, but both often yield low-quality or low-diversity data. We introduce Adversarial Arena for building high quality conversational datasets by framing data generation as an adversarial task: attackers create prompts, and defenders generate responses. This interactive competition between multiple teams naturally produces diverse and complex data. We validated this approach by conducting a competition with 10 academic teams from top US and European universities, each building attacker or defender bots. The competition, focused on safety alignment of LLMs in cybersecurity, generated 19,683 multi-turn conversations. Fine-tuning an open-source model on this dataset produced an 18.47% improvement in secure code generation on CyberSecEval-Instruct and 29.42% improvement on CyberSecEval-MITRE.

Large Language Model (LLM) agents have achieved rapid adoption and demonstrated remarkable capabilities across a wide range of applications. To improve reasoning and task execution, modern LLM agents would incorporate memory modules or retrieval-augmented generation (RAG) mechanisms, enabling them to further leverage prior interactions or external knowledge. However, such a design also introduces a group of critical privacy vulnerabilities: sensitive information stored in memory can be leaked through query-based attacks. Although feasible, existing attacks often achieve only limited performance, with low attack success rates (ASR). In this paper, we propose ADAM, a novel privacy attack that features data distribution estimation of a victim agent's memory and employs an entropy-guided query strategy for maximizing privacy leakage. Extensive experiments demonstrate that our attack substantially outperforms state-of-the-art ones, achieving up to 100% ASRs. These results thus underscore the urgent need for robust privacy-preserving methods for current LLM agents.

Multimodal Large Language Models (MLLMs) have been shown to be vulnerable to malicious queries that can elicit unsafe responses. Recent work uses prompt engineering, response classification, or finetuning to improve MLLM safety. Nevertheless, such approaches are often ineffective against evolving malicious patterns, may require rerunning the query, or demand heavy computational resources. Steering the activations of a frozen model at inference time has recently emerged as a flexible and effective solution. However, existing steering methods for MLLMs typically handle only a narrow set of safety-related concepts or struggle to adjust specific concepts without affecting others. To address these challenges, we introduce Dictionary-Aligned Concept Control (DACO), a framework that utilizes a curated concept dictionary and a Sparse Autoencoder (SAE) to provide granular control over MLLM activations. First, we curate a dictionary of 15,000 multimodal concepts by retrieving over 400,000 caption-image stimuli and summarizing their activations into concept directions. We name the dataset DACO-400K. Second, we show that the curated dictionary can be used to intervene activations via sparse coding. Third, we propose a new steering approach that uses our dictionary to initialize the training of an SAE and automatically annotate the semantics of the SAE atoms for safeguarding MLLMs. Experiments on multiple MLLMs (e.g., QwenVL, LLaVA, InternVL) across safety benchmarks (e.g., MM-SafetyBench, JailBreakV) show that DACO significantly improves MLLM safety while maintaining general-purpose capabilities.

Large language models (LLMs) remain susceptible to jailbreak and direct prompt-injection attacks, yet the strongest defensive filters frequently over-refuse benign queries and degrade user experience. Previous work on jailbreak and prompt injection detection such as GradSafe, detects unsafe prompts with a single "accept all" anchor token, but its threshold is brittle and it offers no deterministic guarantee that harmful content will not be emitted once decoding begins. We introduce Gradient-Controlled Decoding (GCD), a training-free guardrail that combines an acceptance anchor token ("Sure") and refusal anchor token ("Sorry") tightening the decision boundary and significantly lowering false positives. In the mitigation stage, if a prompt is flagged, GCD preset-injects one or two refusal tokens ("Sorry, I can't...") before autoregressive decoding resumes, guaranteeing first-token safety regardless of sampling strategy. On ToxicChat, XSTest-v2, and AdvBench, GCD reduces false positives by 52% vs. GradSafe at comparable recall, lowers attack success rate by up to 10% vs. the strongest decoding-only baseline, adds under 15-20 ms latency on an average on V100 instances, transfers to LLaMA-2-7B, Mixtral-8x7B, and Qwen-2-7B, and requires only 20 demonstration templates.

While prior work has focused on projecting adversarial examples back onto the manifold of natural data to restore safety, we argue that a comprehensive understanding of AI safety requires characterizing the unsafe regions themselves. This paper introduces a framework for systematically mapping the Manifold of Failure in Large Language Models (LLMs). We reframe the search for vulnerabilities as a quality diversity problem, using MAP-Elites to illuminate the continuous topology of these failure regions, which we term behavioral attraction basins. Our quality metric, Alignment Deviation, guides the search towards areas where the model's behavior diverges most from its intended alignment. Across three LLMs: Llama-3-8B, GPT-OSS-20B, and GPT-5-Mini, we show that MAP-Elites achieves up to 63% behavioral coverage, discovers up to 370 distinct vulnerability niches, and reveals dramatically different model-specific topological signatures: Llama-3-8B exhibits a near-universal vulnerability plateau (mean Alignment Deviation 0.93), GPT-OSS-20B shows a fragmented landscape with spatially concentrated basins (mean 0.73), and GPT-5-Mini demonstrates strong robustness with a ceiling at 0.50. Our approach produces interpretable, global maps of each model's safety landscape that no existing attack method (GCG, PAIR, or TAP) can provide, shifting the paradigm from finding discrete failures to understanding their underlying structure.

Large language models (LLMs) are increasingly deployed in security-sensitive applications, where they must follow system- or developer-specified instructions that define the intended task behavior, while completing benign user requests. When adversarial instructions appear in user queries or externally retrieved content, models may override intended logic. Recent defenses rely on supervised fine-tuning with benign and malicious labels. Although these methods achieve high attack rejection rates, we find that they rely on narrow correlations in defense data rather than harmful intent, leading to systematic rejection of safe inputs. We analyze three recurring shortcut behaviors induced by defense fine-tuning. \emph{Position bias} arises when benign content placed later in a prompt is rejected at much higher rates; across reasoning benchmarks, suffix-task rejection rises from below \textbf{10\%} to as high as \textbf{90\%}. \emph{Token trigger bias} occurs when strings common in attack data raise rejection probability even in benign contexts; inserting a single trigger token increases false refusals by up to \textbf{50\%}. \emph{Topic generalization bias} reflects poor generalization beyond the defense data distribution, with defended models suffering test-time accuracy drops of up to \textbf{40\%}. These findings suggest that current prompt-injection defenses frequently respond to attack-like surface patterns rather than the underlying intent. We introduce controlled diagnostic datasets and a systematic evaluation across two base models and multiple defense pipelines, highlighting limitations of supervised fine-tuning for reliable LLM security.

Most jailbreak methods achieve high attack success rates (ASR) but require attacker LLMs to craft adversarial queries and/or demand high query budgets. These resource limitations make jailbreaking expensive, and the queries generated by attacker LLMs often consist of non-interpretable random prefixes. This paper introduces Lexical Anchor Tree Search (), addressing these limitations through an attacker-LLM-free method that operates purely via lexical anchor injection. LATS reformulates jailbreaking as a breadth-first tree search over multi-turn dialogues, where each node incrementally injects missing content words from the attack goal into benign prompts. Evaluations on AdvBench and HarmBench demonstrate that LATS achieves 97-100% ASR on latest GPT, Claude, and Llama models with an average of only ~6.4 queries, compared to 20+ queries required by other methods. These results highlight conversational structure as a potent and under-protected attack surface, while demonstrating superior query efficiency in an era where high ASR is readily achievable. Our code will be released to support reproducibility.

Production LLM agents with tool-using capabilities require security testing despite their safety training. We adapt Go-Explore to evaluate GPT-4o-mini across 28 experimental runs spanning six research questions. We find that random-seed variance dominates algorithmic parameters, yielding an 8x spread in outcomes; single-seed comparisons are unreliable, while multi-seed averaging materially reduces variance in our setup. Reward shaping consistently harms performance, causing exploration collapse in 94% of runs or producing 18 false positives with zero verified attacks. In our environment, simple state signatures outperform complex ones. For comprehensive security testing, ensembles provide attack-type diversity, whereas single agents optimize coverage within a given attack type. Overall, these results suggest that seed variance and targeted domain knowledge can outweigh algorithmic sophistication when testing safety-trained models.

Retrieval-Augmented Generation (RAG) systems enhance LLMs with external knowledge but introduce a critical attack surface: corpus poisoning. While recent studies have demonstrated the potential of such attacks, they typically rely on impractical assumptions, such as white-box access or known user queries, thereby underestimating the difficulty of real-world exploitation. In this paper, we bridge this gap by proposing MIRAGE, a novel multi-stage poisoning pipeline designed for strict black-box and query-agnostic environments. Operating on surrogate model feedback, MIRAGE functions as an automated optimization framework that integrates three key mechanisms: it utilizes persona-driven query synthesis to approximate latent user search distributions, employs semantic anchoring to imperceptibly embed these intents for high retrieval visibility, and leverages an adversarial variant of Test-Time Preference Optimization (TPO) to maximize persuasion. To rigorously evaluate this threat, we construct a new benchmark derived from three long-form, domain-specific datasets. Extensive experiments demonstrate that MIRAGE significantly outperforms existing baselines in both attack efficacy and stealthiness, exhibiting remarkable transferability across diverse retriever-LLM configurations and highlighting the urgent need for robust defense strategies.

Large language models remain vulnerable to jailbreak attacks that bypass safety guardrails to elicit harmful outputs. Defending against novel jailbreaks represents a critical challenge in AI safety. Adversarial training -- designed to make models robust against worst-case perturbations -- has been the dominant paradigm for adversarial robustness. However, due to optimization challenges and difficulties in defining realistic threat models, adversarial training methods often fail on newly developed jailbreaks in practice. This paper proposes a new paradigm for improving robustness against unseen jailbreaks, centered on the Adversarial Déjà Vu hypothesis: novel jailbreaks are not fundamentally new, but largely recombinations of adversarial skills from previous attacks. We study this hypothesis through a large-scale analysis of 32 attack papers published over two years. Using an automated pipeline, we extract and compress adversarial skills into a sparse dictionary of primitives, with LLMs generating human-readable descriptions. Our analysis reveals that unseen attacks can be effectively explained as sparse compositions of earlier skills, with explanatory power increasing monotonically as skill coverage grows. Guided by this insight, we introduce Adversarial Skill Compositional Training (ASCoT), which trains on diverse compositions of skill primitives rather than isolated attack instances. ASCoT substantially improves robustness to unseen attacks, including multi-turn jailbreaks, while maintaining low over-refusal rates. We also demonstrate that expanding adversarial skill coverage, not just data scale, is key to defending against novel attacks. \textcolor{red}{\textbf{Warning: This paper contains content that may be harmful or offensive in nature.

Large language model (LLM) based search agents iteratively generate queries, retrieve external information, and reason to answer open-domain questions. While researchers have primarily focused on improving their utility, their safety behaviors remain underexplored. In this paper, we first evaluate search agents using red-teaming datasets and find that they are more likely to produce harmful outputs than base LLMs. For example, when asked "How can I track someone's location without their consent?", a base model refuses, whereas a search agent designed to retrieve and cite sources may lower its refusal threshold, fetch documents (e.g., court cases), and, once appended, synthesize them into an informative yet unsafe summary. We further show that utility-oriented fine-tuning intensifies this risk, motivating joint alignment of safety and utility. We present SafeSearch, a multi-objective reinforcement learning approach that couples a final-output safety/utility reward with a novel query-level shaping term that penalizes unsafe queries and rewards safe ones. Experiments show that SafeSearch reduces agent harmfulness by over 70% across three red-teaming datasets while producing safe, helpful responses, and matches the QA performance of a utility-only finetuned agent; further analyses confirm the effectiveness of the query-level reward in jointly improving safety and utility.

Recent advances in large reasoning models (LRMs) have enabled remarkable performance on complex tasks such as mathematics and coding by generating long Chain-of-Thought (CoT) traces. In this paper, we identify and systematically analyze a critical vulnerability we term reasoning distraction, where LRMs are diverted from their primary objective by irrelevant yet complex tasks maliciously embedded in the prompt. Through a comprehensive study across diverse models and benchmarks, we show that even state-of-the-art LRMs are highly susceptible, with injected distractors reducing task accuracy by up to 60%. We further reveal that certain alignment techniques can amplify this weakness and that models may exhibit covert compliance, following hidden adversarial instructions in reasoning while concealing them in the final output. To mitigate these risks, we propose a training-based defense that combines Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) on synthetic adversarial data, improving robustness by over 50 points on challenging distractor attacks. Our findings establish reasoning distraction as a distinct and urgent threat to LRM reliability and provide a practical step toward safer and more trustworthy reasoning systems.

Large Language Models (LLMs) can produce catastrophic responses in conversational settings that pose serious risks to public safety and security. Existing evaluations often fail to fully reveal these vulnerabilities because they rely on fixed attack prompt sequences, lack statistical guarantees, and do not scale to the vast space of multi-turn conversations. In this work, we propose C$^3$LLM, a novel, principled statistical Certification framework for Catastrophic risks in multi-turn Conversation for LLMs that bounds the probability of an LLM generating catastrophic responses under multi-turn conversation distributions with statistical guarantees. We model multi-turn conversations as probability distributions over query sequences, represented by a Markov process on a query graph whose edges encode semantic similarity to capture realistic conversational flow, and quantify catastrophic risks using confidence intervals. We define several inexpensive and practical distributions--random node, graph path, and adaptive with rejection. Our results demonstrate that these distributions can reveal substantial catastrophic risks in frontier models, with certified lower bounds as high as 70% for the worst model, highlighting the urgent need for improved safety training strategies in frontier LLMs.

The rapid advancement of Large Language Models (LLMs) has enabled the generation of highly realistic synthetic data. We identify a new vulnerability, LLMs generating convincing career trajectories in fake resumes and explore effective detection methods. To address this challenge, we construct a dataset of machine-generated career trajectories using LLMs and various methods, and demonstrate that conventional text-based detectors perform poorly on structured career data. We propose CareerScape, a novel heterogeneous, hierarchical multi-layer graph framework that models career entities and their relations in a unified global graph built from genuine resumes. Unlike conventional classifiers that treat each instance independently, CareerScape employs a structure-aware framework that augments user-specific subgraphs with trusted neighborhood information from a global graph, enabling the model to capture both global structural patterns and local inconsistencies indicative of synthetic career paths. Experimental results show that CareerScape outperforms state-of-the-art baselines by 5.8-85.0% relatively, highlighting the importance of structure-aware detection for machine-generated content.

AI systems for software development are rapidly gaining prominence, yet significant challenges remain in ensuring their safety. To address this, Amazon launched the Trusted AI track of the Amazon Nova AI Challenge, a global competition among 10 university teams to drive advances in secure AI. In the challenge, five teams focus on developing automated red teaming bots, while the other five create safe AI assistants. This challenge provides teams with a unique platform to evaluate automated red-teaming and safety alignment methods through head-to-head adversarial tournaments where red teams have multi-turn conversations with the competing AI coding assistants to test their safety alignment. Along with this, the challenge provides teams with a feed of high quality annotated data to fuel iterative improvement. Throughout the challenge, teams developed state-of-the-art techniques, introducing novel approaches in reasoning-based safety alignment, robust model guardrails, multi-turn jail-breaking, and efficient probing of large language models (LLMs). To support these efforts, the Amazon Nova AI Challenge team made substantial scientific and engineering investments, including building a custom baseline coding specialist model for the challenge from scratch, developing a tournament orchestration service, and creating an evaluation harness. This paper outlines the advancements made by university teams and the Amazon Nova AI Challenge team in addressing the safety challenges of AI for software development, highlighting this collaborative effort to raise the bar for AI safety.

Large Language Models (LLMs) fine-tuned to align with human values often exhibit alignment drift, producing unsafe or policy-violating completions when exposed to adversarial prompts, decoding perturbations, or paraphrased jailbreaks. While prior work has behaviorally characterized alignment failure, little is known about the training-time belief sources underlying these failures. We introduce TraceAlign, a unified framework for tracing unsafe completions back to their root causes in the model's training corpus. Central to our approach is the Belief Conflict Index (BCI), which quantifies semantic inconsistency between generated spans and aligned policies, based on retrieved training documents using suffix-array matching. We propose three complementary interventions: (i) TraceShield, an inference-time safety filter that refuses completions with high-BCI spans, (ii) Contrastive Belief Deconfliction Loss, a contrastive fine-tuning objective penalizing high-BCI continuations during DPO, and (iii) Prov-Decode, a provenance-aware decoding strategy that vetoes beam expansions predicted to yield high-BCI spans. Together, these defenses reduce alignment drift by up to 85% on our curated Alignment Drift Benchmark (ADB) while preserving utility on standard tasks, with delta less than 0.2 and improved refusal quality. We further derive a theoretical upper bound on drift likelihood via suffix-array span statistics, linking memorization frequency and length to adversarial reactivation risk. TraceAlign thus provides the first scalable, traceable, and grounded toolkit for understanding and mitigating alignment failures at source. To encourage further exploration and development, we open-source our implementation at: https://anonymous.4open.science/r/tracealign-2DA7