arXiv cs.CL
·
18 hours ago
Researchers introduced Code-MUE, a black-box framework that measures uncertainty in code-generating language models by analyzing runtime behavior through Semantic Interaction Graphs rather than text similarity. The method achieved Spearman's correlation of up to -0.98 with functional correctness across eight state-of-the-art code LLMs, substantially outperforming text-based and embedding-based alternatives. This enables better risk detection for deploying code models in production, where distinguishing confident predictions from stochastic guessing is critical for safety and security.
arXiv cs.CL
·
18 hours ago
Researchers developed a graph-based framework to detect disinformation narratives spreading across Russian and Ukrainian Telegram channels by combining weak supervision with propagation graph analysis. The method groups semantically related claims into narrative clusters and models their diffusion across interconnected channels to identify coordinated amplification patterns. This approach enables detection of disinformation spread at the narrative level rather than individual posts, providing insights into how false information propagates through large messaging networks.
arXiv cs.AI
·
18 hours ago
Researchers developed an explainable AI system to detect conversational scams that span multiple weeks or months by analyzing conversation history rather than isolated messages. The system achieved 97.8% accuracy on ConScamBench-278, a new benchmark dataset with 278 examples across eight scam types, and users reported significantly increased confidence in identifying suspicious conversations after using it. The approach addresses limitations of existing single-message phishing detectors by tracking how scammers gradually build trust before requesting money or sensitive information.
arXiv cs.AI
·
18 hours ago
Researchers proposed a policy learning technique using imitation learning to help autonomous cyber-defense agents predict attacker actions in partially observable network environments. The method integrates with neurosymbolic agents using behavior trees with learning-enabled components to defend networks while maintaining operations. The approach achieves high prediction accuracy across simulated scenarios with different attacker policies.
arXiv cs.AI
·
18 hours ago
Researchers demonstrated a cross-application attack called Action Rebinding that allows malicious Android apps with zero permissions to hijack GUI agents powered by large multimodal models and perform privileged operations like deleting files or sending SMS. The attack exploits the observation-action gap in the agent's reasoning pipeline by rendering a benign app to trigger an intended action, then swapping to a sensitive target app during reasoning latency, with a 100% success rate across six tested Android GUI agents. This attack bypasses Android's application sandboxing security model and evades malware detection tools because the malicious app contains no privileged API calls, creating a fundamental security conflict between GUI automation agents and mobile platform security architecture.
arXiv cs.AI
·
18 hours ago
Researchers developed Traffic-Aware Randomized Smoothing (TA-RS), a certified defense method for LLM-based intrusion detection systems that adds Gaussian noise only to network features attackers can control during both training and certification. The method achieved 55-100% certified accuracy on CIC-IDS-2018 and HIKARI-2021 datasets at sigma=0.25, with certified radii exceeding the baseline threshold by 1.8-5 times. This approach improves robustness against traffic manipulation attacks by aligning the noise injection with the attacker's actual capabilities rather than applying uniform noise across all features.
arXiv cs.AI
·
18 hours ago
Researchers surveyed 21 proposals for permission systems in AI agents and analyzed five commercial agents to understand how user-level permissions are specified, derived, and enforced. The study created a taxonomy of permissions handling across user interfaces and internal system design. The work identifies gaps between academic proposals and commercial implementations, indicating areas where agent permission systems need further development to prevent unauthorized actions and data leaks.
arXiv cs.AI
·
18 hours ago
A framework combining Isolation Forest anomaly detection with SHAP explanations was developed to identify fraudulent banking transactions while providing auditors with feature-level justifications for flagged transactions. The model achieved 0.91 precision and 0.88 recall on synthetic banking data, outperforming three unsupervised baselines. The explainability layer measurably improved auditor confidence and decision quality in compliance workflows, enabling practical deployment of transparent AI in regulated financial environments.
arXiv cs.AI
·
18 hours ago
Researchers developed an Adversarial Prompting Framework to test AI safety by generating structured attacks of varying complexity against generative AI models. The framework includes direct harmful requests and advanced encoding-based attacks, with encoded prompts showing the highest success rates at bypassing safety mechanisms. Organizations can now use automated testing to identify vulnerabilities across different attack vectors and quantify their AI systems' security resilience.
arXiv cs.AI
·
18 hours ago
A systematic review analyzed 34 research studies on AI applications in cyberpsychology, which combines psychology with cybersecurity to analyze behavioral patterns of victims, attackers, and defenders. The studies were categorized across four cybersecurity applications—Anomaly Detection, Vulnerability Risk Prediction, Security Awareness Training, and Authentication/Identity Verification—using AI methods including machine learning, deep learning, natural language processing, and reinforcement learning. The review identified psychological concepts used, quantified datasets employed, and highlighted research gaps and emerging methodologies in the AI-cyberpsychology field.
