Academic papers, new techniques, benchmarks, and theoretical findings in AI/LLM security.
Research analyzing 1.5 million ChatGPT messages from workers shows that when employees use AI for tasks outside their normal job description (a pattern called task crossover), some of these activities become regular parts of their work over time rather than one-time experiments. Workers approach cross-occupation tasks differently, writing shorter prompts and asking AI to verify information rather than teach them entirely new skills, suggesting they use AI to borrow expertise from other fields.
ATLAS v2026.09 is an updated database of AI security threats that now includes 120 techniques and 40 mitigations for defending AI systems. The update adds new attack methods like prompt injection (tricking an AI by hiding malicious instructions in its input), AI agent compromises, and exposed AI infrastructure scanning, along with new defensive strategies including AI honeypots (fake AI systems designed to catch attackers).
This research paper analyzes security issues in a blockchain-based system designed to encrypt medical data (using public key encryption with keyword search, which allows searching encrypted data without decrypting it first) when that data is stored in cloud environments. The study examines vulnerabilities in how this encryption system works when combined with blockchain (a distributed ledger technology) for sharing patient information securely.
Researchers analyzed privacy risks in mobile app license agreements and terms of use by creating a dataset of these documents. The study examined what personal data mobile apps claim they can collect and use, revealing potential privacy concerns hidden in the lengthy legal agreements most users don't read. This research helps identify where mobile apps may pose risks to user privacy.
This research paper examines methods that use machine learning (algorithms that learn patterns from data) to identify and trap crypto ransomware (malicious software that encrypts files and demands payment) by analyzing how files behave. The study focuses on evaluating different techniques for selecting which files should be monitored as decoys to detect ransomware attacks before they cause damage.
This research paper evaluates whether large language models (LLMs, AI systems trained on vast amounts of text data) can effectively detect phishing messages (fraudulent messages designed to steal information) on Telegram, a messaging platform. The study examines how well LLMs perform at this security task compared to traditional detection methods.
Researchers discovered a new type of attack called CAFBA (context-aware adaptive fusion backdoor attack) that can compromise AI models used for polyp segmentation (identifying abnormal growths in medical images). This backdoor attack (a hidden malicious instruction planted in an AI model) tricks the medical AI into making incorrect diagnoses when specific conditions are present, potentially causing serious harm to patients.
This academic paper examines privacy concerns in federated learning models (a technique where AI systems train on data spread across multiple locations without centralizing it) used for intrusion detection systems (software that identifies unauthorized access attempts). The research, published in September 2026, appears to focus on understanding how privacy can be protected when building security AI systems across distributed networks.
This research paper introduces AttackLogGen, a benchmark tool that tests how well large language models (LLMs) can generate realistic attack logs (detailed records of suspicious or malicious activity on computer systems). The study evaluates different LLMs' ability to create these logs, which is important for training security systems and testing how well they can detect threats.
This is a survey paper that examines how to make RAG (retrieval-augmented generation, where an AI pulls in external documents to answer questions) more trustworthy when used with large language models. The paper reviews current methods and challenges in ensuring that RAG systems provide reliable and accurate information rather than generating false or misleading answers.
This study examines how teams in virtual environments (like online gaming platforms) are affected when members join or leave in cascades, meaning one person's departure or arrival influences others to do the same. The research found that certain types of team members help keep teams stable during these changes: collaboration-focused members tend to attract new joiners, while arbitration-focused members (those who like coordinating different resources and maintaining control) tend to stay even when others are leaving.
This survey article examines how generative AI (machine learning models that create new content) can produce deepfakes (synthetic media where a person's face or voice is digitally manipulated to appear authentic) and discusses methods for detecting them. The paper reviews the current state of deepfake creation and detection technologies, providing an outlook on future developments in this rapidly evolving field.
This academic survey examines how AI and intelligent technologies affect security in IoV (Internet of Vehicles, where cars and vehicles connect to networks and each other). The paper discusses integrating edge computing (processing data on devices near the source rather than sending everything to distant servers) with AI to improve IoV security, exploring both benefits and challenges in this emerging field.
This is a research survey article that compares security risks across different types of AI systems, including LLMs (large language models, which are AI systems trained on massive amounts of text), AI agents (programs that can make decisions and take actions autonomously), and embodied agents (AI systems that interact with the physical world through robots or similar hardware). The article examines various security threats that affect these different AI systems and how those risks differ between them.
This academic paper describes HPGA, an algorithm designed to anonymize graph data (networks of connected nodes and edges) while preserving personalized information. The research, published in December 2026, addresses the challenge of protecting privacy in graph-structured datasets, which are commonly used in social networks and recommendation systems.
This academic paper presents a new method for detecting intrusions (unauthorized access or attacks) in Industrial Control Systems (ICS, which are computers that manage physical infrastructure like power plants or factories) by combining temporal data (information about when events happen) with symbolic data (categorical information like event types). The approach aims to improve security monitoring in critical infrastructure by better identifying suspicious network activity patterns.
This research proposes a new framework called Πcoll-min for TLS attestation (proving that data came from a secure website connection) in decentralized applications that use blockchain smart contracts. The framework improves upon existing methods by allowing multiple independent verifiers (validators) to jointly check attestations instead of relying on a single designated verifier, which reduces the risk of a prover and verifier secretly cooperating to cheat the system. The authors demonstrate that their approach significantly reduces computational work for the prover (the party providing the attestation) while maintaining security.
Researchers have demonstrated a new attack called Adapdoor that can inject hidden malicious behaviors into Deep Reinforcement Learning (DRL) models, which are AI systems trained to make sequential decisions in environments like robotics and autonomous vehicles. The attack works by poisoning the reward signal (the feedback that guides what the AI learns to do) during training, allowing attackers to later manipulate the model's actions when it is deployed in the real world. The paper shows this threat is more serious than previously thought because Adapdoor can work across many different tasks without requiring manual customization for each one.
Android malware threatens user privacy and data, so researchers use machine learning to detect it, but attackers can craft adversarial malware (malware modified to fool detection systems) that bypasses these defenses. This paper proposes ADD, a defense framework that works as a plug-in to make ML-based malware detection more robust against realistic adversarial attacks, and tests show it effectively protects multiple detection methods and real antivirus solutions.
Intrusion detection systems (IDS, software that monitors networks for unauthorized access attempts) trained on known attack data struggle to detect new, unknown attacks because they have rigid decision boundaries that don't adapt well. This research proposes a new method that uses virtual classes (imaginary attack categories placed between known attacks) and soft labels (flexible probability scores instead of strict categories) to better distinguish between known and unknown attacks, achieving significantly higher detection accuracy on standard network security datasets.