Decision-Level Hijacking: Injecting Cognitive Bias into Large Language Models via Bit-Flip Attacks

Large Language Models (LLMs) have been widely applied in high-stakes decision-making scenarios such as corporate strategy, and users are increasingly relying on their outputs. However, the deep integration of open-source model sharing ecosystems with LLM-powered critical decision-making applications also introduces critical risks: if an attacker can manipulate the model's cognitive stance, they can indirectly influence the judgments and actions of downstream decision-makers. This paper defines such threats as decision-level hijacking. Existing attacks fail to achieve targeted cognitive manipulation without triggering prohibited content or degrading model functionality. To fill this gap, this paper reveals that Bit-Flip Attacks (BFAs) can serve as an attack vector for inducing decision-level hijacking, requiring no real-time interaction or control over the training process, and only a minimal number of weight bits need to be flipped after deployment to achieve stealthy, low-cost, and persistent cognitive manipulation. Therefore, we propose CogBias, a cognitive bias injection framework for LLMs. CogBias converts subjective preferences into optimization signals via a differentiable sentiment evaluator, uses a multi-objective loss to jointly constrain multiple dimensions, and constructs BitScout to locate critical bits, achieving targeted cognitive intervention under an ultra-sparse flip budget. Experiments on Llama-3.2-3B, Mistral-7B, and Qwen2.5-14B, as well as on the commercial recommendation and controversial factual topic scenarios, demonstrate that flipping only a small number of bits stably induces significant stance shifts on target topics, while the impact on non-target tasks and overall output distribution is limited. This work demonstrates that minute perturbations to low-level weight data suffice to undermine the high-level value alignment of LLMs.

Lilith: Backdoor Generalization under Training-Inference Trigger Shift

Machine-learning services increasingly rely on public data, third-party providers, and outsourced training, creating opportunities for data-poisoning attacks that implant persistent malicious behavior while preserving benign utility. However, existing backdoor studies largely evaluate exact trigger reuse, training-exposed trigger diversity, or variations along predefined transformation axes. They therefore leave a critical blind spot: whether a backdoor learned from one training-time trigger can generalize to an inference-time trigger family absent from victim training. We formulate this problem as backdoor generalization under training--inference trigger shift and introduce Lilith, a black-box anchor-to-family framework. Using only disjoint surrogate resources, Lilith first induces a compact target-side vulnerability with a single training anchor, then constructs a bounded inference-only family that preserves the anchor-induced representation geometry. We characterize this mechanism through anchor clearance and family reach, deriving sufficient conditions for family-wise target preservation under local regularity and bounded surrogate--victim discrepancy. Experiments across datasets, architectures, poisoning rates, and defenses show that Lilith achieves high family-wise attack success with limited utility degradation and a small trigger generalization gap. Additional analyses show that family activation depends on representation alignment rather than the proposal mechanism, exposing a broader threat overlooked by exact-trigger evaluation.

Compiling Activation Steering into Weights via Null-Space Constraints for Stealthy Backdoors

Safety-aligned large language models (LLMs) are increasingly deployed in real-world pipelines, yet this deployment also enlarges the supply-chain attack surface: adversaries can distribute backdoored checkpoints that behave normally under standard evaluation but jailbreak when a hidden trigger is present. Recent post-hoc weight-editing methods offer an efficient approach to injecting such backdoors by directly modifying model weights to map a trigger to an attacker-specified response. However, existing methods typically optimize a token-level mapping that forces an affirmative prefix (e.g., ``Sure''), which does not guarantee sustained harmful output -- the model may begin with apparent agreement yet revert to safety-aligned refusal within a few decoding steps. We address this reliability gap by shifting the backdoor objective from surface tokens to internal representations. We extract a steering vector that captures the difference between compliant and refusal behaviors, and compile it into a persistent weight modification that activates only when the trigger is present. To preserve stealthiness and benign utility, we impose a null-space constraint so that the injected edit remains dormant on clean inputs. The method is efficient, requiring only a small set of examples and admitting a closed-form solution. Across multiple safety-aligned LLMs and jailbreak benchmarks, our method achieves high triggered attack success while maintaining non-triggered safety and general utility.

ACIArena: Toward Unified Evaluation for Agent Cascading Injection

Collaboration and information sharing empower Multi-Agent Systems (MAS) but also introduce a critical security risk known as Agent Cascading Injection (ACI). In such attacks, a compromised agent exploits inter-agent trust to propagate malicious instructions, causing cascading failures across the system. However, existing studies consider only limited attack strategies and simplified MAS settings, limiting their generalizability and comprehensive evaluation. To bridge this gap, we introduce ACIArena, a unified framework for evaluating the robustness of MAS. ACIArena offers systematic evaluation suites spanning multiple attack surfaces (i.e., external inputs, agent profiles, inter-agent messages) and attack objectives (i.e., instruction hijacking, task disruption, information exfiltration). Specifically, ACIArena establishes a unified specification that jointly supports MAS construction and attack-defense modules. It covers six widely used MAS implementations and provides a benchmark of 1,356 test cases for systematically evaluating MAS robustness. Our benchmarking results show that evaluating MAS robustness solely through topology is insufficient; robust MAS require deliberate role design and controlled interaction patterns. Moreover, defenses developed in simplified environments often fail to transfer to real-world settings; narrowly scoped defenses may even introduce new vulnerabilities. ACIArena aims to provide a solid foundation for advancing deeper exploration of MAS design principles.

"I See What You Did There": Can Large Vision-Language Models Understand Multimodal Puns?

Puns are a common form of rhetorical wordplay that exploits polysemy and phonetic similarity to create humor. In multimodal puns, visual and textual elements synergize to ground the literal sense and evoke the figurative meaning simultaneously. Although Vision-Language Models (VLMs) are widely used in multimodal understanding and generation, their ability to understand puns has not been systematically studied due to a scarcity of rigorous benchmarks. To address this, we first propose a multimodal pun generation pipeline. We then introduce MultiPun, a dataset comprising diverse types of puns alongside adversarial non-pun distractors. Our evaluation reveals that most models struggle to distinguish genuine puns from these distractors. Moreover, we propose both prompt-level and model-level strategies to enhance pun comprehension, with an average improvement of 16.5% in F1 scores. Our findings provide valuable insights for developing future VLMs that master the subtleties of human-like humor via cross-modal reasoning.

When Agents "Misremember" Collectively: Exploring the Mandela Effect in LLM-based Multi-Agent Systems

Recent advancements in large language models (LLMs) have significantly enhanced the capabilities of collaborative multi-agent systems, enabling them to address complex challenges. However, within these multi-agent systems, the susceptibility of agents to collective cognitive biases remains an underexplored issue. A compelling example is the Mandela effect, a phenomenon where groups collectively misremember past events as a result of false details reinforced through social influence and internalized misinformation. This vulnerability limits our understanding of memory bias in multi-agent systems and raises ethical concerns about the potential spread of misinformation. In this paper, we conduct a comprehensive study on the Mandela effect in LLM-based multi-agent systems, focusing on its existence, causing factors, and mitigation strategies. We propose MANBENCH, a novel benchmark designed to evaluate agent behaviors across four common task types that are susceptible to the Mandela effect, using five interaction protocols that vary in agent roles and memory timescales. We evaluate agents powered by several LLMs on MANBENCH to quantify the Mandela effect and analyze how different factors affect it. Moreover, we propose strategies to mitigate this effect, including prompt-level defenses (e.g., cognitive anchoring and source scrutiny) and model-level alignment-based defense, achieving an average 74.40% reduction in the Mandela effect compared to the baseline. Our findings provide valuable insights for developing more resilient and ethically aligned collaborative multi-agent systems.

FraudShield: Knowledge Graph Empowered Defense for LLMs against Fraud Attacks

Large language models (LLMs) have been widely integrated into critical automated workflows, including contract review and job application processes. However, LLMs are susceptible to manipulation by fraudulent information, which can lead to harmful outcomes. Although advanced defense methods have been developed to address this issue, they often exhibit limitations in effectiveness, interpretability, and generalizability, particularly when applied to LLM-based applications. To address these challenges, we introduce FraudShield, a novel framework designed to protect LLMs from fraudulent content by leveraging a comprehensive analysis of fraud tactics. Specifically, FraudShield constructs and refines a fraud tactic-keyword knowledge graph to capture high-confidence associations between suspicious text and fraud techniques. The structured knowledge graph augments the original input by highlighting keywords and providing supporting evidence, guiding the LLM toward more secure responses. Extensive experiments show that FraudShield consistently outperforms state-of-the-art defenses across four mainstream LLMs and five representative fraud types, while also offering interpretable clues for the model's generations.

Bridging the Copyright Gap: Do Large Vision-Language Models Recognize and Respect Copyrighted Content?

Large vision-language models (LVLMs) have achieved remarkable advancements in multimodal reasoning tasks. However, their widespread accessibility raises critical concerns about potential copyright infringement. Will LVLMs accurately recognize and comply with copyright regulations when encountering copyrighted content (i.e., user input, retrieved documents) in the context? Failure to comply with copyright regulations may lead to serious legal and ethical consequences, particularly when LVLMs generate responses based on copyrighted materials (e.g., retrieved book experts, news reports). In this paper, we present a comprehensive evaluation of various LVLMs, examining how they handle copyrighted content -- such as book excerpts, news articles, music lyrics, and code documentation when they are presented as visual inputs. To systematically measure copyright compliance, we introduce a large-scale benchmark dataset comprising 50,000 multimodal query-content pairs designed to evaluate how effectively LVLMs handle queries that could lead to copyright infringement. Given that real-world copyrighted content may or may not include a copyright notice, the dataset includes query-content pairs in two distinct scenarios: with and without a copyright notice. For the former, we extensively cover four types of copyright notices to account for different cases. Our evaluation reveals that even state-of-the-art closed-source LVLMs exhibit significant deficiencies in recognizing and respecting the copyrighted content, even when presented with the copyright notice. To solve this limitation, we introduce a novel tool-augmented defense framework for copyright compliance, which reduces infringement risks in all scenarios. Our findings underscore the importance of developing copyright-aware LVLMs to ensure the responsible and lawful use of copyrighted content.

The Eminence in Shadow: Exploiting Feature Boundary Ambiguity for Robust Backdoor Attacks

Deep neural networks (DNNs) underpin critical applications yet remain vulnerable to backdoor attacks, typically reliant on heuristic brute-force methods. Despite significant empirical advancements in backdoor research, the lack of rigorous theoretical analysis limits understanding of underlying mechanisms, constraining attack predictability and adaptability. Therefore, we provide a theoretical analysis targeting backdoor attacks, focusing on how sparse decision boundaries enable disproportionate model manipulation. Based on this finding, we derive a closed-form, ambiguous boundary region, wherein negligible relabeled samples induce substantial misclassification. Influence function analysis further quantifies significant parameter shifts caused by these margin samples, with minimal impact on clean accuracy, formally grounding why such low poison rates suffice for efficacious attacks. Leveraging these insights, we propose Eminence, an explainable and robust black-box backdoor framework with provable theoretical guarantees and inherent stealth properties. Eminence optimizes a universal, visually subtle trigger that strategically exploits vulnerable decision boundaries and effectively achieves robust misclassification with exceptionally low poison rates (< 0.1%, compared to SOTA methods typically requiring > 1%). Comprehensive experiments validate our theoretical discussions and demonstrate the effectiveness of Eminence, confirming an exponential relationship between margin poisoning and adversarial boundary manipulation. Eminence maintains > 90% attack success rate, exhibits negligible clean-accuracy loss, and demonstrates high transferability across diverse models, datasets and scenarios.

NeuroBreak: Unveil Internal Jailbreak Mechanisms in Large Language Models

In deployment and application, large language models (LLMs) typically undergo safety alignment to prevent illegal and unethical outputs. However, the continuous advancement of jailbreak attack techniques, designed to bypass safety mechanisms with adversarial prompts, has placed increasing pressure on the security defenses of LLMs. Strengthening resistance to jailbreak attacks requires an in-depth understanding of the security mechanisms and vulnerabilities of LLMs. However, the vast number of parameters and complex structure of LLMs make analyzing security weaknesses from an internal perspective a challenging task. This paper presents NeuroBreak, a top-down jailbreak analysis system designed to analyze neuron-level safety mechanisms and mitigate vulnerabilities. We carefully design system requirements through collaboration with three experts in the field of AI security. The system provides a comprehensive analysis of various jailbreak attack methods. By incorporating layer-wise representation probing analysis, NeuroBreak offers a novel perspective on the model's decision-making process throughout its generation steps. Furthermore, the system supports the analysis of critical neurons from both semantic and functional perspectives, facilitating a deeper exploration of security mechanisms. We conduct quantitative evaluations and case studies to verify the effectiveness of our system, offering mechanistic insights for developing next-generation defense strategies against evolving jailbreak attacks.