Large Language Model Safety

The rapid advancement of large language models (LLMs) has made machine-generated text increasingly difficult to distinguish from human-written text. While recent studies explore leveraging internal representations of language models to uncover deeper detection signals, these raw features often exhibit substantial overlap between classes, limiting their discriminative power. To address this challenge, we propose Steer-to-Detect (\texttt{S2D}), a two-stage framework for detecting LLM-generated text. In the first stage, \texttt{S2D} learns a steering vector that is injected into the hidden states of a frozen observer LLM, producing representations with improved class separability. In the second stage, detection is performed via a hypothesis testing procedure based on the steered representations. We establish finite-sample, high-probability guarantees for Type I and Type II errors, providing a theoretical characterization of the procedure. Empirically, \texttt{S2D} achieves strong and consistent performance across a range of settings, including out-of-distribution scenarios and adversarial perturbations.

Large language models (LLMs) are increasingly deployed in a wide range of applications, yet remain vulnerable to adversarial jailbreak attacks that circumvent their safety guardrails. Existing evaluation frameworks typically report binary success/failure metrics, failing to capture the temporal dynamics of how attacks succeed under persistent adversarial pressure. This preliminary work proposes a novel evaluation framework that applies survival analysis techniques to characterize LLM jailbreak vuln`erability. Our approach models the time-to-jailbreak as a survival outcome, enabling estimation of hazard functions, survival curves, and risk factors associated with successful attacks. We evaluate three LLMs against a subset of prompts from the HarmBench dataset spanning three attack categories. Our analysis reveals that models exhibit distinct vulnerability profiles: while one model demonstrates rapid degradation under iterative attacks, the two other models show consistent moderate vulnerability. Our framework provides actionable insights for model and LLM application developers and establishes survival analysis as a rigorous methodology for LLM safety evaluation.

Large language models (LLMs) achieve strong performance across many tasks but remain vulnerable to hallucinations, making it important to systematically evaluate their reliability under realistic adversarial inputs. We formulate hallucination elicitation as a constrained optimization problem, where the goal is to find semantically coherent adversarial prompts that are equivalent to benign user prompts. Existing attack methods remain limited: discrete prompt-based attacks preserve semantic equivalence and coherence but search only over a limited set of prompt variations, while continuous latent-space attacks explore a richer space but often decode into prompts that are no longer valid rephrasings. To address these limitations, we propose REALISTA, a realistic latent-space attack framework. REALISTA constructs an input-dependent dictionary of valid editing directions, each corresponding to a semantically equivalent and coherent rephrasing, and optimizes continuous combinations of these directions in latent space. This design combines the optimization flexibility of continuous attacks with the semantic realism of discrete rephrasing-based attacks. Experiments demonstrate that REALISTA achieves superior or comparable performance to state-of-the-art realistic attacks on open-source LLMs and, crucially, succeeds in attacking large reasoning models under free-form response settings, where prior realistic attacks fail. Code is available at https://github.com/Buyun-Liang/REALISTA.

We investigate frameworks for post-flight safety analysis at non-towered airports using large language models (LLMs). Non-towered airports rely on the Common Traffic Advisory Frequency (CTAF) for air traffic coordination and experience frequent near mid-air collisions due to the pilot self-announcement communication protocol. We propose a general vision-language model (VLM) approach to analyze the transcribed CTAF radio communications in natural language, METeorological Aerodrome Report (METAR) weather data, Automatic Dependent Surveillance-Broadcast (ADS-B) flight trajectories, and Visual Flight Rules sectional charts of the airfield. We provide a preliminary study at Half Moon Bay Airport, with a qualitative real world case study and a quantitative evaluation using a new synthetic dataset of communications and weather modalities. We qualitatively evaluate our framework on real flight data using Gemini 2.5 Pro, demonstrating accurate identification of a right-of-way violation. The synthetic dataset is derived from real examples and includes a 12-category hazard taxonomy, and is used to benchmark three open-source (Qwen 2.5-7B, Mistral-7B, Gemma-2-9B) and three closed-source (GPT-4o, GPT-5.4, Claude Sonnet 4.6) LLM models on the subset of inputs related to CTAF and METAR. Even limited to CTAF and METAR inputs and open source LLMs, instances of our framework typically achieve a macro F1 score above 0.85 on a binary nominal/danger classification task. Future work includes a quantitative evaluation across all modalities and a larger number of real world examples. Taken together, our results suggest that VLM analysis of safety at non-towered airports may be a valuable future capability.

Supervised Finetuning (SFT) has become one of the primary methods for adapting a large language model (LLM) with extensive pre-trained knowledge to domain-specific, instruction-following tasks. SFT datasets, composed of instruction-response pairs, often include user-provided information that may contain sensitive data such as personally identifiable information (PII), raising privacy concerns. This paper studies the problem of PII reconstruction from SFT models for the first time. We construct multi-turn, user-centric Q&A datasets in sensitive domains, specifically medical and legal settings, that incorporate PII to enable realistic evaluation of leakage. Using these datasets, we evaluate the extent to which an adversary, with varying levels of knowledge about the fine-tuning dataset, can infer sensitive information about individuals whose data was used during SFT. In the reconstruction setting, we propose COVA, a novel decoding algorithm to reconstruct PII under prefix-based attacks, consistently outperforming existing extraction methods. Our results show that even partial attacker knowledge can significantly improve reconstruction success, while leakage varies substantially across PII types.

Conversational scams, such as romance and investment scams, are emerging as a major form of online fraud. Unlike one-shot scam lures such as fake lottery or unpaid toll messages, they unfold through multi-turn conversations in which scammers gradually manipulate victims using evolving psychological techniques. However, existing research mainly focuses on static scam detection or synthetic scams, leaving open whether language models can understand how real-world scams progress over time. We introduce PreScam, a benchmark for modeling scam progression from early conversations. Built from user-submitted scam reports, PreScam filters and structures 177,989 raw reports into 11,573 conversational scam instances spanning 20 scam categories. Each instance is hierarchically structured according to the scam lifecycle defined by the proposed scam kill chain, and further annotated at the turn level with scammer psychological actions and victim responses. We benchmark models on two tasks: real-time termination prediction, which estimates whether a conversation is approaching the termination stage, and scammer action prediction, which forecasts the scammer's subsequent actions. Results show a clear gap between surface-level fluency and progression modeling: supervised encoders substantially outperform zero-shot LLMs on real-time termination prediction, while next-action prediction remains only moderately successful even for strong LLMs. Taken together, these results show that current models can capture some scam-related cues, yet still struggle to track how risk escalates and how manipulation unfolds across turns.

Automated red-teaming for LLMs often discovers narrow attack slices, missing diverse real-world threats, and yielding insufficient data for safety fine-tuning. We introduce Persona-Conditioned Adversarial Prompting (PCAP), which conditions adversarial search on diverse attacker personas (e.g., doctors, students, malicious actors) and strategy sets to explore realistic attack scenarios. By running parallel persona-conditioned searches, PCAP discovers transferable jailbreaks across different contexts and generates rich defense datasets with automatic metadata tracking. On GPT-OSS 120B, PCAP increases attack success from 57% to 97% while producing 2-6$\times$ more diverse prompts covering varied real-world scenarios. Critically, fine-tuning lightweight adapters on PCAP-generated data significantly improves model robustness (recall: 0.36 $\rightarrow$ 0.99, F1: 0.53 $\rightarrow$ 0.96) with minimal false positives, demonstrating a practical closed-loop approach from vulnerability discovery to automated alignment.

Aligning large language models (LLMs) with human values typically relies on post-training or inference-time steering that directly manipulates the backbone's parameters or representation space. However, a critical gap exists: the model's residual stream is highly dynamic, in which values exist as fragile, low-dimensional properties, inherently incompatible with the stability required for consistent value expression. In this paper, we propose the Stable Value Guidance Transformer (SVGT), which addresses this gap through an independent value module incorporating two key designs: (1) independent value modeling, maintaining normative representations in a dedicated value space isolated from the backbone, and (2) explicit behavioral guidance, transducing these stable signals into learnable latent Bridge Tokens. These tokens serve as dynamic value anchors to explicitly steer the generative trajectory, ensuring robust adherence across diverse contexts without disrupting the backbone's internal representations. Experiments across multiple backbones and safety benchmarks show that SVGT generally reduces harmful scores by over 70% while maintaining generation fluency, demonstrating the efficacy of architecturally grounded value modeling. Our code is available at https://github.com/Clervils/SVGT.git.

Backdoor vulnerabilities widely exist in the fine-tuning of large language models(LLMs). Most backdoor poisoning methods operate mainly at the token level and lack deeper semantic manipulation, which limits stealthiness. In addition, Prior attacks rely on a single fixed trigger to induce harmful outputs. Such static triggers are easy to detect, and clean fine-tuning can weaken the trigger-target association. Through causal validation, we observe that emotion is not directly linked to individual words, but functions as an overall stylistic factor through tone. In the representation space of LLM, emotion can be decoupled from semantics, forming distinct cluster from the original neutral text. Therefore, we consider the emotional factor as the backdoor trigger to propose a pparasitic emotion-style dynamic backdoor attack, Paraesthesia. By mixing samples with the emotional trigger into clean data and then fine-tuning the model, the model is able to generate the predefined attack response when encountering emotional inputs during the inference stage. Paraesthesia includes two the quantification and rewriting of emotional styles. We evaluate the effectiveness of our method on instruction-following generation and classification tasks. The experimental results show that Paraesthesia achieves an attack success rate of around 99% across both task types and four different models, while maintaining the clean utility of the models.

Recent published evidence from frontier laboratories shows that contemporary AI models can recognise evaluation contexts, latently represent them, and behave differently under those contexts than under deployment-continuous conditions. Anthropic's BrowseComp incident, the Natural Language Autoencoder findings on SWE-bench Verified and destructive-coding evaluations, and the OpenAI / Apollo anti-scheming work all document instances of this phenomenon. We argue that these findings create a claim-validity problem for safety conclusions drawn from frontier evaluations. We introduce the Evaluation Differential (ED), a conditional divergence in a target behavioural property between recognised-evaluation and deployment-continuous contexts, define a normalised effect-size form (nED) for cross-property comparison, and prove that marginal evaluation scores cannot identify ED. We develop a typology of safety claims (ED-stable, ED-degraded, ED-inverted, ED-undetermined) by their warrant-status under documented divergence, and specify TRACE (Test-Recognition Audit for Claim Evaluation), an audit protocol that wraps existing evaluation infrastructure and produces restricted claims rather than capability scores. We apply the framework retrospectively to three publicly documented evaluation incidents and discuss governance implications for system cards, conformity assessment, and the international network of AI safety and security institutes. TRACE does not eliminate adversarial adaptation; it disciplines the claims drawn from evaluation evidence by making explicit the conditions under which that evidence was produced.

Large Language Model (LLM) agents have emerged as key intermediaries, orchestrating complex interactions between human users and a wide range of digital services and LLM infrastructures. While prior research has extensively examined the security of LLMs and agents in isolation, the systemic risk of the agent acting as a disruptive hub within the user-agent-service chain remains largely overlooked. In this work, we expose a novel threat paradigm by introducing Mobius Injection, a sophisticated attack that weaponizes autonomous agents into zombie nodes to launch what we define as gent-based and -Oriented DDoS (AbO-DDoS) attacks. By exploiting a structural vulnerability in agentic logic named Semantic Closure, an adversary can induce sustained recursive execution of agent components through a single textual injection. We demonstrate that this attack is exceptionally lightweight, stealthy against both traditional DDoS monitors and contemporary AI safety filters, and highly configurable, allowing for surgical targeting of specific environments or model providers. To evaluate the real-world impact, we conduct extensive experiments across three representative claw-style agents and three mainstream coding agents, integrated with 12 frontier proprietary or open-weight LLMs. Our results demonstrate that Mobius Injection achieves substantial attack success across diverse tasks, driving single-node call amplification up to 51.0x and multi-node p95 latency inflation up to 229.1x. The attack performance exhibits a superlinear increase with the number of poisoning nodes. To mitigate Mobius Injection, we propose a proactive defense mechanism using Agent Component Energy (ACE) Analysis, which detects malicious recursive triggers by measuring anomalous energy in the agent's component graph.

The cosine similarity between a large language model's hidden activations before and after Supervised Fine-Tuning (SFT) remains very high. This, at first glance, suggests that SFT leaves the model's activation geometry largely undisturbed. However, projecting both sets of activations through a Sparse Autoencoder (SAE) pretrained on the base model reveals that the underlying sparse latents diverge significantly. We introduce a novel investigative pipeline which utilizes these pretrained SAEs as a high-resolution diagnostic tool to mechanistically investigate the drivers of this representational divergence. Through our analytical pipeline, we discover task-specific and layer-specific distributions of the precise semantic features that are systematically altered during supervised fine-tuning. We additionally identify a layer-wise update profile specific to safety alignment. All code, experimental scripts, and analysis files associated with this work are publicly available at: https://github.com/ruhzi/sae-investigation.

Recent frontier large language models (LLMs) have shown strong performance in identifying security vulnerabilities in large, mature open-source systems. As LLM-generated code becomes increasingly common, a natural goal is to prevent such models from producing vulnerable implementations in the first place. Formal verification offers a principled route to this objective, but existing verification pipelines typically require specifications written in rigid formal languages. Prior work has explored using LLMs to synthesize such specifications, with limited success. In this paper, we investigate a different approach: using LLMs both to generate specifications and to verify implementations compositionally when the specifications are expressed in natural language. Our preliminary results suggest that this approach is promising.

SQL injection (SQLi) attacks are still one of the serious attacks ranked in the Open Worldwide Application Security Project (OWASP) Top 10 threats. Today, with advances in Artificial Intelligence (AI), especially in Large Language Models (LLMs), an opportunity has been created for automating adversarial attack tests to measure the defense mechanisms. In this paper, we aim to create a comprehensive evaluation of use cases that utilize LLMs for adversarial SQL injection generation. We introduce two novel LLM-based systems, Retrieval Augmented Generation for Adversarial SQLi (RADAGAS) and Reflective Chain-of-Thought SQLi (RefleXQLi), and compare them with existing baselines against 10 Web Application Firewalls (WAFs) and one execution-based MySQL validator. To perform a comprehensive test, we used six rule-based open-source WAFs (ModSecurity PL1--3, Coraza PL1--3), 2 AI/ML-based WAFs (WAF Brain, CNN-WAF), and 2 commercial WAFs (AWS WAF and Cloudflare WAF). For the LLM models, we used GPT-4o, Claude 3.7 Sonnet, and DeepSeek R1. Our tests consist of 240 experiments that generate 240,000 payloads and perform 2.2 million tests against WAFs. Our comprehensive evaluation reveals that RADAGAS-GPT4o outperforms other baseline models with a 22.73% bypass rate. The proposed RADAGAS variants are highly successful on AI/ML-based WAFs (92.49% on WAF-Brain by RADAGAS-DeepSeek, 80.48% on CNN-WAF by RADAGAS-Claude), but struggle to bypass rule-based WAFs (0--5.70% on ModSecurity and Coraza). In addition to these findings, another observation is that creating less diverse payloads achieves more bypasses, however they show poor results if the initially chosen payload is not successful. We observe that our findings provide a comprehensive view on using LLM-based approaches in security testing.

Guardrail Classifiers defend production language models against harmful behavior, but although results seem promising in testing, they provide no formal guarantees. Providing formal guarantees for such models is hard because "harmful behavior" has no natural specification in a discrete input space: and the standard epsilon-ball properties used in other domains do not carry semantic meaning. We close this gap by shifting verification from the discrete input space to the classifier's pre-activation space, where we define a harmful region as a convex shape enclosing the representations of known harmful prompts. Because the sigmoid classification head is monotonic, certifying the worst-case point is sufficient to certify the entire region, yielding a closed-form soundness proof without approximation in O(d) time. To formally evaluate these classifiers, we propose two constructions of such regions: SVD-aligned hyper-rectangles, which yield exact SAT/UNSAT certificates, and Gaussian Mixture Models, which yield probabilistic certificates over semantically coherent clusters. Applying this framework to three author-trained Guardrail Classifiers on the toxicity domain, every hyper-rectangle configuration returns SAT, exposing verifiable safety holes across all classifiers, despite seemingly high empirical metrics. Probabilistic GMM certificates also expose a divergent structural stability in how these models represent harm. While GPT-2 and Llama-3.1-8B maintain robust coverage of 90% and 80% across varying boundaries, BERT's safety guarantees prove uniquely volatile. This 'coverage collapse' to 55% at the optimal threshold reveals a sparsely populated safety margin in BERT, which only achieves full coverage by adopting an extremely conservative pessimistic threshold. These approaches combined, provide new insights on how effective Guardrail Classifiers really are, beyond traditional red-teaming.

This paper demonstrates RUBEN, an interactive tool for discovering minimal rules to explain the outputs of retrieval-augmented large language models (LLMs) in data-driven applications. We leverage novel pruning strategies to efficiently identify a minimal set of rules that subsume all others. We further demonstrate novel applications of these rules for LLM safety, specifically to test the resiliency of safety training and effectiveness of adversarial prompt injections.

AI pentesting agents are increasingly credible as offensive security systems, but current benchmarks still provide limited guidance on which will perform best in real-world targets. Existing evaluation protocols assess and optimize for predefined goals such as capture-the-flag, remote code execution, exploit reproduction, or trajectory similarity, in simplified or narrow settings. These tools are valuable for measuring bounded capabilities, yet they do not adequately capture the complexity, open-ended exploration, and strategic decision-making required in realistic pentesting. In this paper, we present a practical evaluation protocol that shifts assessment from task completion to validated vulnerability discovery, allowing evaluation in sufficiently complex targets spanning multiple attack surfaces and vulnerability classes. The protocol combines structured ground-truth with LLM-based semantic matching to identify vulnerabilities, bipartite resolution to score findings under realistic ambiguity, continuous ground-truth maintenance, repeated and cumulative evaluation of stochastic agents, efficiency metrics, and reduced-suite selection for sustainable experimentation. This protocol extends the state of the art by enabling a more realistic, operationally informative comparison of AI pentesting agents. To enable reproducibility, we also release expert-annotated ground truth and code for the proposed evaluation protocol: https://github.com/ethiack/ethibench.

The rapid proliferation of LLM-based autonomous agents in real operating system environments introduces a new category of safety risk beyond content safety: behavior jailbreak, where an adversary induces an agent to execute dangerous OS-level operations with irreversible consequences. Existing benchmarks either evaluate safety at the semantic layer alone, missing physical-layer harms, or fail to isolate test cases, letting earlier runs contaminate later ones. We present LITMUS (LLM-agents In-OS Testing for Measuring Unsafe Subversion), a benchmark addressing both gaps via a semantic-physical dual verification mechanism and OS-level state rollback. LITMUS comprises 819 high-risk test cases organized into one harmful seed subset and six attack-extended subsets covering three adversarial paradigms (jailbreak speaking, skill injection, and entity wrapping), plus a fully automated multi-agent evaluation framework judging behavior at both conversational and OS-level physical layers. Evaluation across frontier agents reveals three findings: (1) current agents lack effective safety awareness, with strong models (e.g., Claude Sonnet 4.6) still executing 40.64% of high-risk operations; (2) agents exhibit pervasive Execution Hallucination (EH), verbally refusing a request while the dangerous operation has already completed at the system level, invisible to every prior semantic-only framework; and (3) skill injection and entity wrapping attacks achieve high success rates, exposing pronounced agent vulnerabilities. LITMUS provides the first standardized platform for reproducible, physically grounded behavioral safety evaluation of LLM agents in real OS environments.

LLMs are increasingly deployed as autonomous agents with access to tools, databases, and external services, yet practitioners (across different sectors) lack systematic methods to assess how known threat classes translate into concrete risks within a specific agentic deployment. We present MATRA, a pragmatic threat modeling framework for agentic AI systems that adapts established risk assessment methodology to systematically assess how known LLM threats translate into deployment-specific risks. MATRA begins with an asset-based impact assessment and utilizes attack trees to determine the likelihood of these impacts occurring within the system architecture. We demonstrate MATRA on a personal AI agent deployment using OpenClaw, quantifying how architectural controls such as network sandboxing and least-privilege access reduce risk by limiting the blast radius of successful injections.

The Model Context Protocol (MCP) has become a widely adopted interface for LLM agents to invoke external tools, yet learned monitoring of MCP tool-call traffic remains underexplored. In this article, the proposed detector is presented as an attack detection framework for MCP tool-call traffic that encodes each agent session as a graph (tool calls as nodes, sequential and data-flow links as edges), enriches nodes with sentence-embedding features over arguments and responses, and classifies sessions as benign or attacked. Three GNN architectures (GAT, GCN, GraphSAGE), a no-graph MLP, and classical baselines (XGBoost, random forest, logistic regression, linear SVM) are evaluated, with the full architecture comparison conducted on RAS-Eval (task-stratified splits) and GraphSAGE retained as the GNN baseline on ATBench and a combined-source variant (both label-stratified). Three findings emerge. First, content-level features are essential: metadata-only detection plateaus around an AUROC of 0.64 regardless of architecture, while content embeddings push the AUROC above 0.89. Second, naive random-split evaluation inflates AUROC by up to 26 percentage points relative to task-disjoint splits, a memorization confound that prior agent-detection work has not addressed. Third, the detection signal resides primarily in the SBERT content embeddings: an AUROC of 0.975 was reached by tree ensembles on pooled embeddings, performing, for the most part, better than the neural architectures in the primary RAS-Eval setting including GNNs (0.917) and the MLP (0.896), and self-supervised pre-training does not deliver a label-efficiency advantage on this task.

Fine-tuning Large Language Models (LLMs) on benign narrow data can sometimes induce broad harmful behaviors, a vulnerability termed emergent misalignment (EM). While prior work links these failures to specific directions in the activation space, their relationship to the model's broader persona remains unexplored. We map the latent personality space of LLMs through established psychometric profiles like the Big Five, Dark Triad, and LLM-specific behaviors (e.g. evil, sycophancy), and show that the semantic geometry is highly stable across aligned models and their corrupted fine-tunes. Through causal interventions, we find that directions isolating social valence, such as the 'Evil' persona vector, and a Semantic Valence Vector (SVV) that we introduce, function as intrinsic guardrails: ablating them drives the misalignment rates above $40$%, while amplifying them suppresses the failure mode to less than $3$%. Leveraging the structural stability of the personality space, we show that vectors extracted $\textit{a priori}$ from an instruct-tuned model transfer zero-shot to successfully regulate EM in corrupted fine-tunes. Overall, our findings suggest that harmful fine-tuning does not overwrite a model's internal representation of personality, allowing conserved representations to serve as robust, cross-distribution guardrails.

Multi-agent LLM systems introduce a security risk in which sensitive information accessed by one agent can propagate through shared context and reappear in downstream outputs, even without explicit adversarial intent. We formalise this phenomenon as propagation amplification, where leakage risk increases across agent boundaries as sensitive content is repeatedly exposed to downstream generators. Existing defences, including prompt-based safeguards, static pattern matching, and LLM-as-judge filtering, are not designed for this setting: they either operate after generation, rely primarily on surface-form patterns, or add substantial latency without modelling the generation process itself. To resolve these issues, we propose PRISM, a real-time defence that treats credential leakage as a sequential risk accumulation problem during generation. At each decoding step, PRISM combines 16 signals spanning lexical, structural, information-theoretic, behavioural, and contextual features into a calibrated risk score, enabling per-token intervention through green, yellow, and red risk zones. Our central observation is that credential reproduction is often preceded by a measurable shift in generation dynamics, characterised by entropy collapse and increasing logit concentration. When combined with text-structural cues such as identifier-pattern detection, these temporal signals provide an early warning of leakage before a secret is fully reconstructed. Across a 2,000-task adversarial benchmark covering 13 attack categories and three pressure levels in a heterogeneous four-agent pipeline, PRISM achieves F1 = 0.832 with precision = 1.000 and recall = 0.712, while producing no observed leakage on our benchmark (0.0% task-level leak rate) and preserving output utility of 0.893. It substantially outperforms the strongest baseline, Span Tagger, which achieves F1 = 0.719 with a 15.0% task-level leak rate.

Threat detection in Natural Language Processing lacks consistent definitions and standardized benchmarks, and is often conflated with broader phenomena such as toxicity, hate speech, or offensive language. In this work, we introduce ThreatCore, a public available benchmark dataset for fine-grained threat detection that distinguishes between explicit threats, implicit threats, and non-threats. The dataset is constructed by aggregating multiple publicly available resources and systematically re-annotating them under a unified operational definition of threat, revealing substantial inconsistencies across existing labels. To improve the coverage of underrepresented cases, particularly implicit threats, we further augment the dataset with synthetic examples, which are manually validated using the same annotation protocol adopted for the re-annotation of the public datasets, ensuring consistency across all data sources. We evaluate Perspective API, zero-shot classifiers, and recent language models on ThreatCore, showing that implicit threats remain substantially harder to detect than explicit ones. Our results also indicate that incorporating Semantic Role Labeling as an intermediate representation can improve performance by making the structure of harmful intent more explicit. Overall, ThreatCore provides a more consistent benchmark for studying fine-grained threat detection and highlights the challenges that current models still face in identifying indirect expressions of harmful intent.

AI scientist systems are increasingly deployed for autonomous research, yet their academic integrity has never been systematically evaluated. We introduce SCIINTEGRITY-BENCH, the first benchmark designed around a dilemmatic evaluation paradigm: each of its 33 scenarios across 11 trap categories is constructed so that honest acknowledgment of failure is the only correct response, while task completion requires misconduct. Across 231 evaluation runs spanning 7 state-of-the-art LLMs, the overall integrity problem rate reaches 34.2%, and no model achieves zero failures. Most strikingly, across missing-data scenarios, all seven models generate synthetic data rather than acknowledging infeasibility, differing only in whether they disclose the substitution. A further prompt ablation study separates two drivers: removing explicit completion pressure sharply reduces undisclosed fabrication from 20.6% to 3.2%, while the underlying synthesis rate remains unchanged, revealing an intrinsic completion bias that persists independent of prompt-level instructions. These findings point to the absence of honest refusal as a trained disposition as the primary driver of observed failures. We release SCIINTEGRITY-BENCH at https://github.com/liuxingtong/Sci-Integrity-Bench.

Large language models (LLMs) increasingly rely on knowledge editing to support knowledge-intensive reasoning, but this flexibility also introduces critical safety risks: adversaries can inject malicious or misleading knowledge that corrupts downstream reasoning and leads to harmful outcomes. Existing knowledge editing benchmarks primarily focus on editing efficacy and lack a unified framework for systematically evaluating the safety implications of edited knowledge on reasoning behavior. To address this gap, we present EditRisk-Bench, a benchmark for systematically evaluating safety risks of knowledge-intensive reasoning under malicious knowledge editing. Unlike prior benchmarks that mainly emphasize edit success, generalization, and locality, EditRisk-Bench focuses on how injected knowledge affects downstream reasoning behavior and reliability. It integrates diverse malicious scenarios, including misinformation, bias, and safety violations, together with multi-level knowledge-intensive reasoning tasks and representative editing strategies within a unified evaluation framework measuring attack effectiveness, reasoning correctness, and side effects. Extensive experiments on both open-source and closed-source LLMs show that malicious knowledge editing can reliably induce incorrect or unsafe reasoning while largely preserving general capabilities, making such risks difficult to detect. We further identify several key factors influencing these risks, including edit scale, knowledge characteristics, and reasoning complexity. EditRisk-Bench provides an extensible testbed for understanding and mitigating safety risks in knowledge editing for LLMs.

An attacker can split a malicious goal into sub-prompts that each look benign on their own and only become harmful in combination. Existing LLM safety benchmarks evaluate prompts one at a time, or across turns of a single chat, and so do not look for a malicious signal spread across separate sessions with no shared context. We build FragBench, a benchmark drawn from 24 real-world cyber-incident campaigns, which keeps the full attack trail: the multi-fragment kill chain, the per-fragment safety-judge verdicts, sandboxed execution traces, and a matched set of benign cover sessions. FragBench splits this trail into two paired tasks: an adversarial rewriter that hardens fragments against a single-turn safety judge (FragBench Attack), and a graph-based user-level detector trained on the resulting interactions (FragBench Defense). The single-turn judge is near chance on the released corpus by construction, but four GNN variants and three classical-ML baselines all recover the cross-session feature, reaching aggregate event-level F1 = 0.88-0.96. Defending against fragmented LLM misuse therefore requires modeling the cross-session interaction graph, rather than isolated prompts. Our generator, rewriter, sandbox harness, and detector are released at https://github.com/LidaSafety/fragbench.

Defenses against indirect prompt injection (IPI) in tool-using LLM agents share two structural weaknesses. First, they all attempt to prevent attacks rather than detect the compromises that slip through. Second, they have only been evaluated in English, leaving users of low-resource languages such as Kurdish and Arabic without tested protection. This paper addresses both gaps with AgentShield, a deception-based detection framework that places three layers of traps inside the agent's tool interface: fake tools, fake credentials, and allowlisted parameters. The same trap triggers serve as high-precision labels for a self-supervised classifier. An LLM agent that follows an attacker's hidden instruction almost always touches one of these traps, which gives both a real-time compromise signal and a zero-FP label for training a downstream detector without manual annotation. Across 176 cross-lingual attack prompts and four LLMs from three providers, and because modern LLMs already refuse most IPI attempts on their own (attack success rate <= 10%), AgentShield's job is to catch the attacks that do slip through. On commercial models, it catches 90.7%-100% of such successful attacks, with zero false alarms on 485 normal-use tests. It survives a systematic adaptive-attack evaluation with zero evasion on commercial models, and the self-supervised classifier transfers across models and languages without retraining.

Steering or intervening on model representations at inference time to correct predictions is essential for AI interpretability and safety, yet existing evaluation protocols are limited to ambiguous language modeling tasks. To address this gap, we introduce SwordBench, a benchmark for steering image representations of vision models across multiple backbones and concept removal tasks. Beyond a unified benchmarking suite, we propose new evaluation notions that uncover the second-order effects of orthogonalization among concept activation vectors for pragmatic steering. Specifically, cross-concept robustness measures the stability of concept detection performance across inputs orthogonalized against alternative concepts, and collateral damage quantifies whether steering inadvertently affects model performance on a downstream task for inputs lacking the bias. We find that although a linear support vector machine exhibits superior separability and orthogonality, it fails to achieve zero collateral damage, often trailing sparse autoencoders. In simpler regimes, both standard baselines and optimization-based methods fail to achieve perfect steering. The source code will be made available soon on GitHub.

Smart-contract exploits have caused billions of USD in cumulative losses, yet audits remain expensive and slow. Automated tools have emerged to close this gap, but each class has a characteristic failure mode. Static analyzers report findings that frequently fail manual triage at high rates, while large language models (LLMs) hallucinate findings that contradict the source code. Thus, we propose Chaintrix, an end-to-end auditing framework whose central architectural commitment is that every LLM-generated claim must be discharged against a deterministic structural contract representation. We introduce a Cross-Contract Interaction Model (CCIM) that parses Solidity into a structured map of function-level reads, writes, modifiers and resolved cross-contract calls. CCIM serves as the substrate against which all 12 of Chaintrix's deterministic signal engines and the parallel LLM audit pipelines operate. A staged false-positive-reduction pipeline, terminating in a Structural Verdict Engine (SVE) that applies deterministic structural checks against parsed code, filters the merged finding set, with selected high-confidence findings further validated through symbolic execution and fuzz testing. We evaluate Chaintrix on EVMbench, the smart-contract security benchmark by OpenAI, Paradigm, OtterSec. Chaintrix detects 86 of 120 high-severity vulnerabilities (71.7% recall), with 25 audits scoring 100% recall, placing Chaintrix 26 percentage points above the strongest frontier-model baseline.

Jailbreak attacks -- adversarial prompts that bypass LLM alignment through purely linguistic manipulation -- pose a growing operational security threat, yet the field lacks large-scale, reproducible infrastructure for generating, categorizing, and evaluating them systematically. This paper addresses that gap with three contributions. (1) Large-scale compositional jailbreak dataset. We construct 114,000 adversarial prompts by applying 912 composing strategies to 125 harmful seed prompts from JailBreakV-28K. Every prompt is assigned to one of 14 cybersecurity attack categories (e.g., malware, phishing, privilege escalation) via a six-model majority-vote pipeline, and each strategy is ranked by effectiveness per category, enabling principled strategy selection grounded in concrete adversarial objectives. (2) Automated jailbreak generation. We instruction-fine-tune category-aware LLMs on Moderate and Optimal subsets, producing models that synthesize fluent jailbreak prompts from a harmful seed at inference time -- no templates, no gradient search. Our generators achieve perplexity 24-39 versus 40-140 for AutoDAN and AmpleGCG, with safety-filter evasion rates of 0.29-0.51 Mal (LlamaPromptGuard-2-86M), enabling controllable, scalable red-teaming under realistic adversarial conditions. (3) OPTIMUS: a training-free jailbreak evaluator. OPTIMUS is a continuous metric J(S,H) that jointly captures semantic similarity between the harmful seed and the jailbreak (S) and harmfulness probability (H) via calibrated penalty functions. Unlike binary attack success rate (ASR), OPTIMUS requires no task-specific training, generalizes across evolving strategies, and exposes a stealth-optimal regime (S*=0.57, H*=0.43) that ASR misses. Experiments across 114,000 prompts confirm that OPTIMUS separates Weak, Moderate, and Optimal jailbreaks with category-level evidence binary evaluation cannot supply.

In multi-agent systems (MAS), a single deceptive agent can nullify all gains of an agentic AI collective and evade deployed defenses. However, existing adversarial studies on MAS target only shallow tasks and do not consider adaptive adversaries, which evolve their strategies to evade the very detectors trained to catch them. To address that gap, we introduce GAMBIT, a benchmark with three evaluation modes and two independent scores for evaluating imposter detectors: the first two modes measure zero-shot detection under increasing distribution shift, and a third recalibration mode measures how quickly a detector adapts to novel attacks from just 20 labeled examples. The benchmark comes with a dataset of 27,804 labeled instances spanning 240 co-evolved imposter strategies. Our contributions are threefold: (1) Using chess as a substrate deep reasoning problem and Gemini 3.1 Pro for agents, we release GAMBIT and its dataset to evaluate imposter detectors under realistic constraints against a stealthy adaptive imposter; (2) We introduce an adaptive imposter agent based on an efficient evolutionary framework, generalizable beyond chess, that collapses collective task performance while remaining essentially undetectable (50.5% F1-score with a Gemini-based detector); (3) We show that zero-shot evaluation can be highly misleading for adaptive adversaries: two detectors with near-identical zero-shot scores differ by 8x on few-shot adaptation, while the meta-learned variant converges 20x faster, a gap only visible in the recalibration mode. Altogether, GAMBIT provides the first multi-agent benchmark where adversarial attacks and defenses co-evolve, with an imposter framework generalizable beyond our use case, and promising techniques for fast recalibration in a rapidly evolving adversarial system. Code and data: https://anonymous.4open.science/r/gambit.

Large Language Models (LLMs) increasingly employ alignment techniques to prevent harmful outputs. Despite these safeguards, attackers can circumvent them by crafting adversarial prompts. Predominant token-level optimization methods primarily rely on optimizing for exact affirmative templates (e.g., ``\textit{Sure, here is...}''). However, these paradigms frequently encounter bottlenecks such as suboptimal convergence, compromised prompt naturalness, and poor cross-model generalization. To address these limitations, we propose Semantic Representation Attack (SRA), a novel LLM-agnostic paradigm that fundamentally reconceptualizes adversarial objectives from exact textual targeting to malicious semantic representations. Theoretically, we establish the semantic Coherence-Convergence Relationship and derive a Cross-Model Semantic Generalization bound, proving that maintaining semantic coherence guarantees both white-box semantic convergence and black-box transferability. Technically, we operationalize this framework via the Semantic Representation Heuristic Search (SRHS) algorithm, which preserves interpretability and structural coherence of the adversarial prompts during incremental discrete token chunk expansion. Extensive evaluations demonstrate that our framework achieves a 99.71% average attack success rate across 26 open-source LLMs, with strong transferability and stealth.

Large Language Models (LLMs) are increasingly deployed as autonomous agents that execute tool-augmented, multi-step tasks, where latency is a critical factor for real-world applications. Yet an overlooked threat is Reasoning-Level Denial-of-Service (R-DoS), in which an attacker preserves task correctness but degrades availability by inflating an agent's reasoning depth or tool-use budget. We introduce OTora, the first unified, two-stage red-teaming framework for instantiating R-DoS attacks. Stage I optimizes an adversarial trigger that induces targeted tool invocations using insertion-aware scoring and dynamic target co-evolution, supporting both black-box and white-box settings. Stage II generates agent-aware reasoning payloads via an ICL-guided genetic search that amplifies overthinking while maintaining correct task outcomes. Across WebShop, Email, and OS agents built on multiple backbone models such as LLaMA-70B and GPT-OSS-120B, OTora achieves up to 10 times increases in reasoning tokens and order-of-magnitude latency slowdowns, all while preserving near-baseline task accuracy. Finally, we discuss mitigation strategies for detecting and constraining abnormal reasoning and latency spikes. The code is available at https://github.com/llm2409/OTora.

Watermarking for large language models (LLMs) is a promising approach for detecting LLM-generated text and enabling responsible deployment. However, existing watermarking methods are often vulnerable to semantic-invariant attacks, such as paraphrasing. We propose PASA, a principled, robust, and distortion-free watermarking algorithm that embeds and detects a watermark at the semantic level. PASA operates on semantic clusters in a latent embedding space and constructs a distributional dependency between token and auxiliary sequences via shared randomness synchronized by a secret key and semantic history. This design is grounded in our theoretical framework that characterizes a jointly optimal embedding-detection pair, achieving the fundamental trade-offs among detection accuracy, robustness, and distortion. Evaluations across multiple LLMs and semantic-invariant attacks demonstrate that PASA remains robust even under strong paraphrasing attacks while preserving high text quality, outperforming standard vocabulary-space baselines. Ablation studies further validate the effectiveness of our hyperparameter choices. Webpage: https://ai-kunkun.github.io/PASA_page/.

Debugging CUDA programs has long been challenging because failures often arise from subtle interactions among hardware behavior, compiler decisions, memory hierarchy, and asynchronous execution. More importantly, with the rapid expansion of GPU usage across scientific computing, machine learning, graphics, and systems workloads, CUDA debugging has become more challenging than ever. Current evaluations of LLM-based CUDA programming largely miss this setting: a model can pass correctness tests with repair by degeneration, simplifying the CUDA code into a safer but slower program that abandons the original optimization structure. We introduce CUDABEAVER, a benchmark for CUDA debugging from real failing workspaces produced during LLM-based CUDA generation. Each task provides the broken candidate, native build/test commands, raw error evidence, and a single editable file. CUDABEAVER evaluates whether a fixer truly repairs the failing CUDA code or merely finds a slower test-passing replacement, reporting results by failure category, debugging trajectory, stagnation mode, and performance preservation. We further propose pass@k(M,C,A), a protocol-conditional CUDA debugging metric by making the fixer M, corpus C, and protocol axes Aexplicit. Using this metric across 213 tasks and seven frontier LLMs, we show that protocol-aware evaluation gives a more faithful view of CUDA debugging ability: when performance-loss tolerance is high, fixers appear much stronger, but even a minor stricter performance requirement can sharply reduce measured success, shifting scores by up to 40 percentage points.

Ensuring safe, policy-compliant outputs from large language models requires real-time content moderation that can scale across multiple safety dimensions. However, state-of-the-art guardrail models rely on autoregressive decoders with 7B--27B parameters, reformulating what is fundamentally a classification problem as sequential text generation, a design choice that incurs high latency and scales poorly to multi-aspect evaluation. In this work, we introduce \textbf{GLiGuard}, a 0.3B-parameter schema-conditioned bidirectional encoder adapted from GLiNER2 for LLM content moderation. The key idea is to encode task definitions and label semantics directly into the input sequence as structured token schemas, enabling simultaneous evaluation of prompt safety, response safety, refusal detection, 14 fine-grained harm categories, and 11 jailbreak strategies in a single non-autoregressive forward pass. This schema-conditioned design lets supported task and label blocks be composed directly in the input schema at inference time. Across nine established safety benchmarks, GLiGuard achieves F1 scores competitive with 7B--27B decoder-based guards despite being 23--90$\times$ smaller, while delivering up to 16$\times$ higher throughput and 17$\times$ lower latency. These results suggest that compact bidirectional encoders can approach the accuracy of much larger guard models while drastically reducing inference cost. Code and models are available at https://github.com/fastino-ai/GLiGuard.

Federated fine-tuning (FFT) has emerged as a privacy-preserving paradigm for collaboratively adapting large language models (LLMs). Built upon federated learning, FFT enables distributed agents to jointly refine a shared pretrained LLM by aggregating local LLM updates without sharing local raw data. However, FFT-based LLMs remain vulnerable to model manipulation threats, in which adversarial participants upload manipulated LLM updates that corrupt the aggregation process and degrade the performance of the global LLM. In this paper, we propose an Augmented Model maniPulation (AugMP) strategy against FFT-based LLMs. Specifically, we design a novel graph representation learning framework that captures feature correlations among benign LLM updates to guide the generation of malicious updates. To enhance manipulation effectiveness and stealthiness, we develop an iterative manipulation algorithm based on an augmented Lagrangian dual formulation. Through this formulation, malicious updates are optimized to embed adversarial objectives while preserving benign-like parameter characteristics. Experimental results across multiple LLM backbones demonstrate that the AugMP strategy achieves the strongest manipulation performance among all competing baselines, reducing the global LLM accuracy by up to $26\%$ and degrading the average accuracy of local LLM agents by up to $22\%$. Meanwhile, AugMP maintains high statistical and geometric consistency with benign updates, enabling it to evade conventional distance- and similarity-based defense methods.

LLMs are increasingly capable of persuasion, which raises the question of how to protect users against manipulation. In a preregistered user study (N=120) across four decision-making scenarios, we find that an adversarial LLM with a hidden goal succeeds in steering users' decisions 65.4% of the time. We then introduce a "warden" model: a secondary LLM that monitors the human-AI interaction trace in real time and issues non-binding, private advisories to the user when it detects manipulation. Adding a warden more than halves the adversary's success rate to 30.4%, with a much smaller (8.6 percentage points) reduction for genuine interactions. To probe the mechanism behind these results, we release COAX-Bench, a simulation benchmark spanning 14 decision-making scenarios, including hiring, voting, and file access. Across 16,212 simulated multi-agent interactions, capable adversarial LLMs achieve their hidden goals in 34.7% of cases, which warden models reduce to 12.3%. Notably, even warden models substantially weaker than the adversary they oversee provide meaningful protection, suggesting a path for scalable oversight of more capable models.

Conversational Large Language Models are post-trained on language that expresses specific behavioural traits, such as curiosity, open-mindedness, and empathy, and values, such as helpfulness, harmlessness, and honesty. This is done to increase utility, ensure safety, and improve the experience of the people interacting with the model. However, values are complex and inter-related -- inducing one could modify behaviour on another. Further, inducing certain values can make models more addictive or sycophantic through language used in the generations, with a potential detrimental effect on the user. We investigate these and other unintended effects of value induction into models. We fine-tune models using curated value subsets of existing preference datasets, measuring the impact of value induction on expression of other values, model safety, anthropomorphic language, and various QA benchmarks. We find that (i) inducing values leads to expression of other related, and sometimes contrastive values, (ii) inducing positive values increases safety, and (iii) all values increase anthropomorphic language use, making models more validating and sycophantic.

Large Language Models (LLMs) rely on safety alignment to obey safe requests while refusing harmful ones. However, traditional refusal mechanisms often lead to "rigid rejection," where a general template (e.g., "I cannot fulfill this request") indiscriminately triggers refusals and severely undermines the naturalness of interactions between humans and LLMs. To address this issue, LANCE is proposed in this paper to ensure safe yet flexible and natural responses via label enhancement. Specifically, LANCE employs variational inference to perform label enhancement, predicting a continuous distribution across multiple rejection categories. These fine-grained rejection distributions provide multi-way textual gradients for a refinement model to neutralize the hazardous elements in the prompt, so that the LLMs could generate safe responses that avoid rigid rejections while preserving the naturalness of interactions. Experiments demonstrate that LANCE significantly alleviates the rigid rejection problem while maintaining high security standards, significantly outperforming existing baseline models in terms of helpfulness and naturalness of responses.

Large language models (LLMs) rely on deterministic pseudorandom number generators (PRNGs) for autoregressive sampling, creating a critical supply-chain attack surface overlooked by existing defenses. We present SeedHijack, a backdoor attack that manipulates PRNG outputs to force attacker-specified token selection without altering model logits. In a 540-trial benchmark on GPT-2 (124M), the attack achieves 99.6% exact token injection across 9 sampling configurations; it reaches 100% success on four aligned models (1.5B-7B, RLHF/SFT/reasoning distillation) and bypasses all alignment methods tested in this work. We further propose a defense based on a hardware quantum random number generator (QRNG), which neutralizes the attack in our evaluated threat model with negligible median overhead (+0.6% latency, +7.7 MB memory). Our work identifies a critical sampling-layer vulnerability and provides a practical, deployable QRNG-based defense.

Platform content moderation applies explicit policy rules and context-dependent conditions to decide whether user content is allowed, restricted, or removed. A correct moderation outcome must therefore depend on which rules a case activates, how those rules interact, and whether the available evidence is sufficient. Current multimodal safety benchmarks largely reduce moderation to matching predefined final labels, leaving this underlying rule structure untested. As a result, a high benchmark score reveals little about whether a model applies the policy correctly or arrives at the correct label through superficial cues. To evaluate this rule-governed process, we introduce RuleSafe-VL, a benchmark for rule-conditioned decision reasoning in vision-language content moderation. Derived from publicly available platform moderation policies, RuleSafe-VL formalizes 93 atomic rules and 92 typed rule relations, yielding 2,166 context-sensitive image-text cases across three high-risk policy families. Its four diagnostic tasks decompose moderation into a rule-conditioned decision chain. They identify activated rules, recover rule interactions, judge decision sufficiency, and resolve outcomes once missing context is supplied. Experiments on 10 frontier, open-source, and safety-oriented VLMs reveal rule-relation recovery as the dominant bottleneck, where the best model reaches only 64.8 Macro-F1 and some safety-oriented models fall below 7 Macro-F1. Decision-state prediction also remains unreliable, peaking at 64.5 Macro-F1. RuleSafe-VL shifts moderation evaluation from final-label scoring toward diagnostic assessment of rule-conditioned decision reasoning.

In this paper, we investigate the recent state-of-the-art schemes for watermarking large language models (LLMs) outputs. These techniques are claimed to be robust, scalable and production-grade, aimed at promoting responsible usage of LLMs. We analyse the effectiveness of these watermarking techniques against an extensive collection of modified text attacks, which perform targeted semantic changes without altering the general meaning of the text content. Our approach encompasses multiple attack strategies, which include lexical alterations, machine translation, and even neural paraphrasing. The attack efficacy is measured with two target criteria - successful removal of the watermark and preservation of semantic content. We evaluate semantic preservation through BERT scores, text complexity measures, grammatical errors, and Flesch Reading Ease indices. The experimental results reveal varying levels of effectiveness among different watermarking models, with the same underlying result that it is possible to remove the watermark with reasonable effort. This study sheds light on the strengths and weaknesses of existing LLM watermarking systems, suggesting how they should be constructed to improve security of available schemes.

Many-shot jailbreaking (MSJ) causes safety-aligned language models to answer harmful queries by preceding them with many harmful question-answer demonstrations. We study why this attack becomes stronger as the number of demonstrations increases. Empirically, we find that MSJ induces a progressive activation drift: the representation of a fixed harmful query moves step by step away from the safety-aligned region as more harmful demonstrations are added. Theoretically, we show that this drift can be interpreted as implicit malicious fine-tuning: conditioning on N harmful demonstrations induces SGD-style updates equivalent to optimizing on the corresponding N harmful samples. This view turns the attack mechanism into a defense principle. We append a fixed one-shot safety demonstration at inference time, which induces a counteracting safety-oriented update and restores refusal behavior. The resulting method improves the model's robustness to MSJ without modifying its parameters or requiring white-box access at deployment. Code is available at https://github.com/Thecommonirin/SafeEnd.

Multimodal RAG systems increasingly rely on vision-language retrievers to ground visual queries in external textual evidence. Existing adversarial studies on RAG mainly manipulate the retrieval corpus or memory, while attacks on vision-language and remote sensing models typically target end-task predictions. Input-space threats to the evidence retrieval stage of remote sensing multimodal RAG remain underexplored. To address this gap, we introduce CloudWeb, an atmospheric retrieval hijacking attack that modifies only the input image while keeping the retriever, generator, and knowledge base fixed at deployment. CloudWeb overlays parameterized cloud- and haze-like patterns on remote sensing images and optimizes them with a retrieval-oriented objective that pulls adversarial image embeddings toward target atmospheric evidence, suppresses source-scene evidence, enforces rank separation, and regularizes naturalness and coverage. To the best of our knowledge, this is the first study of retrieval-stage atmospheric evidence hijacking in remote sensing multimodal RAG. We evaluate CloudWeb on a seven-dataset remote sensing RAG benchmark with five CLIP-style retrievers, including GeoRSCLIP, RemoteCLIP, OpenAI CLIP, and OpenCLIP, together with downstream vision-language generators. Across retrievers, CloudWeb consistently outperforms clean retrieval, handcrafted atmospheric baselines, random cloud perturbations, and fixed variants in injecting weather-related evidence into top-ranked results. On GeoRSCLIP ViT-B/32, Weather@5 increases from 0.71% to 43.29%. Downstream generation further shows measurable weather hallucination and semantic shift, indicating that retrieval-stage hijacking can propagate to the final RAG response. These findings reveal a practical failure mode: natural-looking atmospheric changes can compromise evidence retrieval before generation begins.

Indirect prompt injection remains a persistent weakness in retrieval-augmented and tool-using LLM systems, and the problem becomes harder to characterise in multilingual settings. We present MIPIAD, a defense framework evaluated on English and Bangla that combines a sequence classifier fine-tuned from Qwen2.5-1.5B via LoRA (XLPID), TF-IDF lexical features, and validation-tuned ensembling through late fusion, stacking, and gradient boosting. The framework is evaluated on a synthetic benchmark built from BIPIA(Yi et al., 2023) templates spanning five task families -- email, table, QA, abstract, and code-comprising over 1.43 million generated samples, with train and test splits using mutually exclusive attack categories. Across the experiments, lexical signals prove strong (TF-IDF+SVM F1=0.77), and the hybrid XLPID+TF-IDF ensemble achieves the best overall F1 (0.9205) while the Boosting Ensemble achieves the best AUROC (0.9378). Ensemble methods consistently reduce the English-Bangla cross-lingual gap relative to standalone neural models. The pipeline is designed for extensibility: NLLB-200 supports over 200 languages and XLPID's multilingual backbone can be retargeted to additional languages without architectural changes; empirical validation is currently limited to English and Bangla

Recent advancements in visual context compression enable MLLMs to process ultra-long contexts efficiently by rendering text into images. However, we identify a critical vulnerability inherent to this paradigm: lowering image resolution inadvertently catalyzes jailbreaking. Our experiments reveal that the safety defenses of SOTA models deteriorate sharply as resolution degrades, surprisingly persisting even when text remains legible. We attribute this to Cognitive Overload'', hypothesizing that the effort required to decipher degraded inputs diverts attentional resources from safety auditing. This phenomenon is consistent across various visual perturbations, including noise and geometric distortion. To address this, we propose a simple Structured Cognitive Offloading'' strategy that mitigates these risks by enforcing a serialized pipeline to decouple visual transcription from safety assessment. Our work exposes a significant risk in vision-based compression and provides critical insights for the secure design of future MLLMs.

A major bottleneck in characterizing the failure modes of generative AI systems is the cost and time of annotation and evaluation. Consequently, adaptive testing paradigms have gained popularity, where one opportunistically decides which cases and how many to annotate based on past results. While this framework is highly practical, its extreme flexibility makes it difficult to draw statistically rigorous conclusions, as it violates classical assumptions: the number of observations is typically limited (often 10 to 50 cases) and decisions regarding sampling and stopping are made in the midst of data collection rather than based a pre-specified rule. To characterize what statistical inferences can be drawn from highly adaptive audits, we introduce a hypothesis testing framework from two 'dueling' perspectives: (i) the model's null that asserts there is no failure mode with performance below a target threshold versus (ii) the auditor's null that asserts they have a sampling strategy that will uncover a failure mode. Leveraging Safe Anytime-Valid Inference (SAVI), we formalize the auditor as conducting 'testing by betting', which translates into simultaneous e-processes for testing the dueling null hypotheses. Furthermore, if the auditor is sufficiently powerful, we prove that these two hypotheses are asymptotically inverses of each other, in that passage of a stringent audit does in fact certify the AI system as being globally robust. Empirically, we demonstrate that our proposed testing procedures maintain anytime-valid type-I error control, outperform pre-specified testing methods, and can reach statistically rigorous conclusions sometimes with as few as 20 observations.

Many deployments must compare candidate language models for safety before a labeled benchmark exists for the relevant language, sector, or regulatory regime. We formalize this setting as benchmarkless comparative safety scoring and specify the contract under which a scenario-based audit can be interpreted as deployment evidence. Scores are valid only under a fixed scenario pack, rubric, auditor, judge, sampling configuration, and rerun budget. Because no labels are available, we replace ground-truth agreement with an instrumental-validity chain: responsiveness to a controlled safe-versus-abliterated contrast, dominance of target-driven variance over auditor and judge artifacts, and stability across reruns. We instantiate the chain in SimpleAudit, a local-first scoring instrument, and validate it on a Norwegian safety pack. Safe and abliterated targets separate with AUROC values between 0.89 and 1.00, target identity is the dominant variance component ($η^2 \approx 0.52$), and severity profiles stabilize by ten reruns. Applying the same chain to Petri shows that it admits both tools. The substantial differences arise upstream of the chain, in claim-contract enforcement and deployment fit. A Norwegian public-sector procurement case comparing Borealis and Gemma 3 demonstrates the resulting evidence in practice: the safer model depends on scenario category and risk measure. Consequently, scores, matched deltas, critical rates, uncertainty, and the auditor and judge used must be reported together rather than collapsed into a single ranking.

Watermark radioactivity testing type of methods can detect whether a model was trained on watermarked documents, and have become key tools for protecting data ownership in the fine-tuning of large language models (LLMs). Existing works have proved their effectiveness in centralized LLM fine-tuning. However, this type of method faces several challenges and remains underexplored in federated learning (FL), a widely-applied paradigm for fine-tuning LLMs collaboratively on private data across different users. FL mainly ensures privacy through secure aggregation (SA), which allows the server to aggregate updates while keeping clients' updates private. This mechanism preserves privacy but makes it difficult to identify which client trained on watermarked documents. In this work, we propose FedAttr, a new client-level attribution protocol for FL. FedAttr identifies which clients trained on watermarked data via a paired-subset-difference mechanism, while preserving the privacy guarantees of SA and FL performance. FedAttr proceeds in three steps: (i) estimate each client's update by differencing two SA queries, (ii) score the estimate with the watermark detector via differential scoring, and (iii) combine scores across rounds via Stouffer method. We theoretically show that FedAttr produces an unbiased estimator of each client's update with bounded mutual information leakage (i.e., $O(d^*/N)$ per-round update). Moreover, FedAttr empirically achieves 100% TPR and 0% FPR, outperforming all baselines by at least 44.4% in TPR or 19.1% in FPR, with only 6.3% overhead relative to FL training time. Ablation studies confirm that FedAttr is robust to protocol parameters and configurations.

This work studies the robust evaluation of iterative stochastic purification defenses under white-box adversarial attacks. Our key technical insight is that gradient checkpointing makes exact end-to-end gradient computation through long purification trajectories practical by trading additional recomputation for substantially lower memory usage. This enables full-gradient adaptive attacks against diffusion- and Langevin-based purification defenses, where prior evaluations often resort to approximate backpropagation due to memory constraints. These approximations can weaken the attack signal and risk overestimating robustness. In parallel, stochasticity in iterative purification is frequently under-controlled, even though different purification trajectories can substantially change reported robustness metrics. Building on this insight, we introduce a memory-efficient full-gradient evaluation framework for stochastic purification defenses. The framework combines checkpointed backpropagation with evaluation protocols that control stochastic variability, thereby reducing memory bottlenecks while preserving exact gradients. We evaluate diffusion-based purification and Langevin sampling with Energy-Based Models (EBMs), demonstrating that full-gradient attacks uncover vulnerabilities missed by approximate-gradient evaluations. Our framework yields stronger state-of-the-art $\ell_{\infty}$ and $\ell_{2}$ white-box attacks and further supports probing out-of-distribution robustness. Overall, our results show that exact-gradient evaluation is essential for reliable benchmarking of iterative stochastic defenses.

Large language models (LLMs) are now largely involved in software development workflows, and the code they generate routinely includes third-party library (TPL) imports annotated with specific version identifiers. These version choices can carry security and compatibility risks, yet they have not been systematically studied. We present the first large-scale measurement study of version-level risk in LLM-generated Python code, evaluating 10 LLMs on PinTrace, a curated benchmark of 1,000 Stack Overflow programming tasks. LLMs tend to specify version identifiers when directly prompted at 26.83%-95.18%, while down to 6.45%-59.19% in creating a manifest file directly. Among the specified versions, 36.70%-55.70% of tasks contain at least one known CVE, and 62.75%-74.51% of them carry Critical or High severity ratings. In 72.27%-91.37% of cases, the associated CVEs were publicly disclosed before the model's knowledge cutoff. The statistics show all models converge on the same small set of risky release versions, indicating a systemic bias rather than isolated model error. Static compatibility rates range from 19.70% to 63.20%, with installation failure as the dominant cause. The dynamic test cases confirm the pattern by 6.49%-48.62% pass rates. Further experiments confirm that these failures are attributable to version selection rather than code quality, and that externally anchored version constraints substantially reduce both vulnerability exposure and compatibility failures. Our findings reveal LLM version selection as a first-class, previously overlooked risk surface in LLM-based development. We disclosed these findings to the community of the evaluated models, and several confirmed the issue. All the code and dataset have been released for open science at https://github.com/dw763j/PinTrace.

LLM-as-a-Judge pipelines have become the de facto evaluator for agent safety, yet existing benchmarks treat their verdicts as ground-truth proxies without checking whether the verdicts depend on the agent's behavior or merely on how the evaluation policy happens to be worded. We argue that any trustworthy safety judge must satisfy a basic property we call policy invariance, and we operationalize it as three testable principles: rubric-semantics invariance under certified-equivalent rewrites, rubric-threshold invariance under intentional strict-to-lenient shifts, and ambiguity-aware calibration so that verdict instability concentrates on genuinely ambiguous cases. Instantiating these principles as a stress-test protocol with four agent-class judges on trajectories drawn from ASSEBench and R-Judge, we surface a previously unmeasured failure mode: today's judges respond to meaningful normative shifts and to meaningless structural rewrites with comparable strength, and cannot tell the two apart. Content-preserving policy rewrites flip up to 9.1% of verdicts above baseline jitter, and 18-43% of all observed flips occur on unambiguous cases under such rewrites, so existing safety scores conflate what the agent did with how the evaluator was prompted. Beyond the diagnosis, we contribute the Policy Invariance Score and the Judge Card reporting protocol, which expose an order-of-magnitude spread in judge reliability that is invisible to accuracy-only leaderboards. We release the protocol and code so that future agent-safety benchmarks can audit their own evaluators rather than trust them by default.

The safety alignment of Large Language Models (LLMs) remains vulnerable to Harmful Fine-tuning (HFT). While existing defenses impose constraints on parameters, gradients, or internal representations, we observe that they can be effectively circumvented under persistent HFT. Our analysis traces this failure to the inherent redundancy of the high-dimensional parameter space: attackers exploit optimization trajectories that are orthogonal to defense constraints to restore harmful capabilities while deceptively adhering to safety restrictions. To address this, we propose Safety Bottleneck Regularization (SBR). SBR shifts the defensive focus from the redundant parameter space to the unembedding layer, which serves as a geometric bottleneck. By anchoring the final hidden states of harmful queries to those of the safety-aligned model, SBR enables the model to maintain safe responses even under persistent HFT. Extensive experiments confirm SBR's effectiveness, demonstrating that utilizing just a single safety anchor is sufficient to reduce the Harmful Score to $<$10 while preserving competitive performance on benign downstream tasks.

Retrieval-Augmented Generation (RAG) enables large language models (LLMs) to leverage external knowledge, but also exposes valuable RAG databases to leakage attacks. As RAG systems grow more complex and LLMs exhibit stronger instruction-following capabilities, existing studies fall short of systematically assessing RAG leakage risks. We present LeakDojo, a configurable framework for controlled evaluation of RAG leakage. Using LeakDojo, we benchmark six existing attacks across fourteen LLMs, four datasets, and diverse RAG systems. Our study reveals that (1) query generation and adversarial instructions contribute independently to leakage, with overall leakage well approximated by their product; (2) stronger instruction-following capability correlates with higher leakage risk; and (3) improvements in RAG faithfulness can introduce increased leakage risk. These findings provide actionable insights for understanding and mitigating RAG leakage in practice. Our codebase is available at https://github.com/yeasen-z/LeakDojo.

Recent advances in foundation models have transformed LLMs from passive conversational systems into autonomous agents capable of reasoning and tool execution. While these capabilities unlock substantial practical value, they also introduce new security risks, as adversaries can manipulate agents into performing harmful actions in real-world environments. Existing defense strategies mitigate such threats but frequently struggle to balance safety and utility, resulting in over-refusal of benign user requests. To mitigate this trade-off, we propose SafeHarbor, a novel framework designed to establish precise decision boundaries for LLM agents. Unlike static guidelines, SafeHarbor extracts context-aware defense rules through enhanced adversarial generation. We design a local hierarchical memory system for dynamic rule injection, offering a training-free, efficient, and plug-and-play solution. Furthermore, we introduce an information entropy-based self-evolution mechanism that continuously optimizes the memory structure through dynamic node splitting and merging. Extensive experiments demonstrate that SafeHarbor achieves state-of-the-art performance on both ambiguous benign tasks and explicit malicious attacks, notably attaining a peak benign utility of 63.6% on GPT-4o while maintaining a robust refusal rate exceeding 93% against harmful requests. The source code is publicly available at https://github.com/ljj-cyber/SafeHarbor.

Recent developments in AI safety research have called for red-teaming methods that effectively surface potential risks posed by generative AI models, with growing emphasis on how red-teamers' backgrounds and perspectives shape their strategies and the risks they uncover. While automated red-teaming approaches promise to complement human red-teaming through larger-scale exploration, existing automated approaches do not account for human identities and rarely incorporate human inputs. In this work, we explore persona-driven red-teaming to advance both automated red-teaming and human-AI collaboration. We first develop PersonaTeaming Workflow, which incorporates personas into the adversarial prompt generation process to explore a wider spectrum of adversarial strategies. Compared to RainbowPlus, a state-of-the-art automated red-teaming method, PersonaTeaming Workflow achieves higher attack success rates while maintaining prompt diversity. However, since automated personas only approximate real human perspectives, we further instantiate PersonaTeaming Workflow as PersonaTeaming Playground, a user-facing interface that enables red-teamers to author their own personas and collaborate with AI to mutate and refine prompts. In a user study with 11 industry practitioners, we found that PersonaTeaming Playground enabled diverse red-teaming strategies and outputs that practitioners perceived as useful, and that AI-generated suggestions in the PersonaTeaming Playground encouraged out-of-the-box thinking even when practitioners did not follow them strictly. Together, our work advances both automated and human-in-the-loop approaches to red-teaming, while shedding light on interaction patterns and design insights for supporting human-AI collaboration in generative AI red-teaming.

Large reasoning models (LRMs) increasingly expose chain-of-thought-like reasoning for transparency, verification, and deliberate problem solving. This creates a safety blind spot: harmful or policy-violating content may appear in reasoning traces even when final answers appear safe. We test whether final-answer safety is a sufficient proxy for the full reasoning-answer trajectory by scoring both stages under a unified twenty-principle safety rubric. Using prompts from seven public harmfulness and jailbreak sources, plus four out-of-distribution (OOD) sources, we evaluate 15 open-weight and API-based LRMs across 41K prompts per model. Reasoning traces consistently reveal additional safety risks beyond final answers, especially in high-severity stage-wise failures: leak cases, where unsafe reasoning precedes a safe-looking answer, and escape cases, where benign-looking reasoning precedes an unsafe final response. Principle-level analysis shows that risk concentrates in misinformation, legal compliance, discrimination, physical harm, and psychological harm. We further propose adaptive multi-principle steering, a white-box test-time mitigation that learns one unsafe-to-safe activation direction per safety principle and activates only directions whose current hidden state is closer to the unsafe than safe centroid. On three steerable open reasoning models, adaptive steering reduces unsafe counts in both reasoning traces and final answers on held-out and OOD benchmarks. DeepSeek-R1-Qwen-7B achieves a 40.8% average unsafe-count reduction while retaining 97.7% macro-averaged accuracy on BBH, GSM8K, and MMLU. These results suggest that LRM safety should be evaluated and mitigated over the full exposed reasoning-answer trajectory, not only at the final-answer stage.

Current LLM safety benchmarks are predominantly English-centric and often rely on translation, failing to capture country-specific harms. Moreover, they rarely evaluate a model's ability to detect culturally embedded sensitivities as distinct from universal harms. We introduce XL-SafetyBench. a suite of 5,500 test cases across 10 country-language pairs, comprising a Jailbreak Benchmark of country-grounded adversarial prompts and a Cultural Benchmark where local sensitivities are embedded within innocuous requests. Each item is constructed via a multi-stage pipeline that combines LLM-assisted discovery, automated validation gates, and dual independent native-speaker annotators per country. To distinguish principled refusal from comprehension failure, we evaluate Attack Success Rate (ASR) alongside two complementary metrics we introduce: Neutral-Safe Rate (NSR) and Cultural Sensitivity Rate (CSR). Evaluating 10 frontier and 27 local LLMs reveals two key findings. First, jailbreak robustness and cultural awareness do not show a coupled relationship among frontier models, so a composite safety score obscures per-axis variation. Second, local models exhibit a near-linear ASR-NSR trade-off (r = -0.81), indicating that their apparent safety reflects generation failure rather than genuine alignment. XL-SafetyBench enables more nuanced, cross-cultural safety evaluation in the multilingual era.

Retrieval-Augmented Generation (RAG) systems are vulnerable to knowledge base poisoning, yet existing attacks have been evaluated almost exclusively against vanilla retrieve-then-generate pipelines. Architectures designed to handle conflicting retrieved information - multi-agent debate, agentic retrieval, recursive language models - remain untested against adversarially optimized contradictions. We evaluate four RAG architectures (vanilla RAG, agentic RAG, MADAM-RAG, and Recursive Language Models) under controlled single-document (N=1) poisoning on 921 Natural Questions QA pairs, comparing a clean baseline, naive injection, and CorruptRAG-AK - an adversarial attack whose meta-epistemic framing targets credibility assessment. Architecture is a high-impact variable in adversarial robustness: under CorruptRAG-AK, attack success rates range from 81.9% (vanilla) to 24.4% (RLM) - a spread of nearly 58 percentage points across architectures with comparable clean accuracy (~92%). Decomposing this gap, once the poisoned document is retrieved, adversarial framing - not retrieval optimization - drives the majority of CorruptRAG-AK's advantage for three of four architectures, localizing the cross-architecture vulnerability at the content-reasoning stage. Our MADAM-RAG reimplementation shows the highest apparent contradiction detection rate, though our LLM judge over-identifies this behavior (~48.5% precision), so reported rates are upper bounds. Regardless of detection, MADAM-RAG cannot resolve contradictions reliably, producing a 41.4% non-answer rate even on clean inputs - though implementation divergences from the original may contribute. We introduce a seven-category behavioral taxonomy capturing contradiction detection, hedging, and failure modes beyond binary accuracy. Code, data, and analysis notebooks are publicly available.