Research

OpenAI discovered that SWE-Bench Pro, a widely-used benchmark for measuring AI coding abilities, has significant quality problems that make it unreliable for evaluating model capabilities. Approximately 30% of the tasks in the benchmark are broken due to issues like overly strict tests, unclear instructions, insufficient test coverage, or misleading prompts, meaning the benchmark no longer accurately measures whether AI models can actually write software.

This research evaluates nine different dataset watermarking methods (techniques for embedding hidden markers into training data to prove ownership of AI models) by testing how well they preserve prediction accuracy, verify ownership, and resist attacks. The study finds that while most watermarking methods cause minimal overall accuracy loss, they can significantly harm performance on specific classes, fail completely when training data is mixed together, and have various other robustness vulnerabilities.

This research paper presents REO (Robust and Efficient 3D semantic Occupancy), a system that helps autonomous vehicles understand their surroundings by combining data from multiple cameras and LiDAR sensors (a laser-based detection system) into a unified 3D map. Unlike existing methods that depend on accurate sensor calibration (precise alignment settings between sensors), REO uses attention mechanisms (a technique that helps AI focus on important parts of data) to learn how to convert 2D camera images into 3D space representations without requiring calibration, making it more practical for real-world driving conditions while running much faster on vehicle computers.

Researchers developed an AI-based attack system that can identify and jam wireless communications across different IoT protocols (Wi-Fi, ZigBee, BLE) operating in the 2.4 GHz frequency band. The system uses deep learning models (LSTM neural networks with attention mechanisms) trained on channel state information (CSI, the detailed characteristics of wireless signals) to classify signal types with over 96% accuracy, then employs reinforcement learning (a type of AI that learns through trial and error) to intelligently decide when and how to jam these communications. Tests showed the attack successfully disrupted ZigBee and BLE device performance while remaining difficult to detect.

This research presents TapPass, an acoustic-based second-factor authentication system (an extra security layer beyond passwords that uses a second verification method) for mobile phones that identifies users by analyzing the unique sound patterns created when they tap their PIN on a device. The system uses deep learning (a type of AI that mimics how brains process information) with special techniques to extract consistent identifying features from these tap sounds and adapt over time, achieving 93% accuracy in tests.

This survey reviews multimodal recommender systems, which are AI models that personalize user experiences by analyzing multiple types of data (such as text, images, and user behavior) together rather than separately. The researchers found that these systems can discover complementary information across different data types that single-type systems might miss, and they provide a framework for implementing and comparing these models.

This research proposes a framework for making AI hiring systems more fair by addressing algorithmic bias (when AI systems make systematically unfair decisions against certain groups). The study analyzes real applicant data and finds that language differences in interviews and how interview questions are structured can cause unfair outcomes, but these problems can be reduced by modifying linguistic features (the words and language patterns used) and making interview questions more consistent across all applicants.

Large Language Models are vulnerable to prompt injection attacks (tricking an AI by hiding malicious instructions in its input), which poses security risks during deployment. This paper proposes BUUAS, a framework that uses Bayesian-inspired belief updates and a belief-weighted contextual multi-armed bandit mechanism (a decision-making approach that learns which security checks to prioritize) to detect prompt injection attacks more efficiently by focusing on high-risk user requests in edge-cloud networks (systems that process data both locally on edge devices and in remote data centers).

Ppin is an anomaly-based intrusion detection system that uses provenance graphs (visual maps showing how different parts of a computer system interact with each other) to identify attacks by finding unusual patterns in system behavior. Unlike existing systems that struggle to explain attacks clearly, Ppin uses a memory-augmented neural network (a machine learning model that remembers information over time) to detect rare behaviors and correlate them to show the full story of how an attack happened.

Environmental illusions like shadows and tire marks can trick autonomous driving systems into misreading lanes, a safety problem that hasn't been studied much until now. Researchers created LanEvil++, a benchmark (a test suite for measuring performance) with 90,000+ images showing 14 types of illusions to evaluate how well lane detection models handle these challenges. The study found that shadows cause the most problems, reducing model accuracy by 5-10%, and proposes the Multimodal Illusion Defense Approach (MIDA, a training method using difficult examples) to improve robustness.

This research analyzed 183 studies to compare how organizations can get employees to follow information security policies (rules about protecting data and systems). The study found that informal controls, like peer influence and self-motivation, are more effective than formal controls like rules and monitoring. Formal controls work indirectly by helping employees develop informal controls through learning and accepting the organization's values.

In early 2024, Google's Gemini AI model generated historically inaccurate and racially offensive images, such as depicting non-White figures in Nazi-era settings, exposing failures in AI training and ethical oversight. The root cause was a flawed "diversity injection" mechanism (a technique meant to reduce bias in training data) that lacked safeguards to understand historical context, resulting in distorted outputs. The incident caused significant financial and reputational damage to Google and raised broader questions about how to build fairness and accuracy into generative AI systems.

This research examines how information systems can build trust in global hydrogen supply chains, which connect energy sources in the Global South and Australia to customers in the Northern Hemisphere. The study identifies five key trust-building techniques: transparency and traceability (making information visible and trackable), economic risk coordination, scenario planning, knowledge sharing, and communication. The findings suggest that information systems can play an important role in making hydrogen a reliable energy source by addressing trust concerns among organizations working together in these emerging supply chains.

AI systems used by organizations often appear reliable but actually suffer from algorithmic fragility, a persistent instability caused by how machine learning interacts with complex real-world environments, leading to biased outputs and performance drift (a decline in accuracy over time). Organizations manage this instability through stabilization work, which involves three practices: buffering (absorbing problems), reframing (reinterpreting failures), and patching (fixing issues), that become routine organizational processes to maintain the appearance of reliability. The paper argues that algorithmic fragility is a structural condition of AI systems rather than a temporary bug, and that effective AI governance requires continuous practice-based management rather than one-time technical fixes.

RAG (retrieval-augmented generation, where an AI pulls in external documents to answer questions) systems can be attacked by poisoning the documents they retrieve, which amplifies biases (systematic unfairness) in AI outputs even when the original question seems neutral. For example, a compromised RAG system might retrieve documents that show women as domestic violence victims, causing the AI to generate biased responses about gender. The paper demonstrates this attack using a framework called BRRA and shows that a dual stage defense mechanism can help reduce these attacks' impact.

This research examines security vulnerabilities in clustered distributed storage systems (CDSSs, which are networks of computers that store data across multiple machines organized into groups), focusing on attacks by two types of adversaries: omniscient adversaries (attackers with complete knowledge of the system) and limited-knowledge adversaries (attackers with partial information). The study analyzes how these attackers can compromise the system by sending corrupted data during file reconstruction and repair, and proposes theoretical bounds on system resilience using coding techniques.

Neural Radiance Fields (NeRF, a technique for representing 3D objects and scenes) need better ways to track which datasets were used to train them, especially since existing watermarking methods require changes during training that don't work well for this purpose. Researchers introduced TraceNeRF, a new method that embeds owner-specific identifying marks directly into training datasets using a hybrid approach combining spatial masking and discrete cosine transform (a mathematical technique for breaking images into frequency components), along with special modules to make these marks harder to remove or alter.

LiDAR subsystems (3D sensing systems that help autonomous vehicles perceive their surroundings) in self-driving cars face network-layer security risks when data packets travel through vehicle networks to the autonomous driving system. This study introduces a framework to trace how these network-level vulnerabilities propagate through the LiDAR data processing pipeline in Apollo V8.0 (a self-driving software platform), identifying three critical processing mechanisms (Frame Partition, Object Erasure, and Object Mark) that determine whether attacks succeed, and showing that failure depends on factors like timing and frame alignment rather than simple data corruption.

Researchers developed PromptFishing, a technique to identify whether social media accounts are controlled by LLMs (large language models, AI systems trained on text) or humans by embedding hidden prompts into normal conversations that cause LLMs to produce distinctive responses called hallucinations (when an AI generates false or off-topic information). The method uses a two-stage algorithm to create adversarial suffixes (specially crafted text additions) that trigger these distinctive LLM behaviors while appearing like normal conversation, achieving over 85% accuracy at detecting LLMs while falsely flagging humans less than 1% of the time.