Research

Researchers show that Kubernetes' lack of visibility into container image downloads, which sit behind the CRI-API, can be exploited for a Denial of Service attack on worker nodes. The attacks reportedly drive average CPU usage up to 95%, block new image downloads, and raise I/O and network usage for a potentially unlimited time. The authors argue a full fix needs a major architectural change to the Kubernetes and CRI-API relationship, which they consider unfeasible in the short term.

The paper proposes Janus, an open and resilient remote attestation scheme for Trusted Execution Environments. It replaces the centralized trust model, which relies on a single provisioned secret key and one verifier, with a Physically Unclonable Function as the root of trust plus decentralized verification on a smart contract. An automated switch mechanism lets the scheme adapt to different situations, and the authors report a UC-based security proof and a complete prototype.

The paper introduces SAGA, an approach for generating fine-grained, labeled synthetic audit logs that mimic real system logs while embedding APT attacks. Malicious events follow the MITRE ATT&CK framework across an APT lifecycle and are blended with benign operational logs over configurable durations. The authors use these datasets to run baseline event-based technique hunting and campaign detection, and show that a deep learning model trained on them can detect previously unseen techniques.

This review surveys AI-based automatic prostate gland segmentation across multi-modal medical imaging, including CT, MRI and ultrasound. It introduces a classification of methods by the quantity and nature of supervision used during training, and assesses the strengths and limitations of existing approaches. The authors also cover evaluation metrics, current challenges and future research directions.

The authors argue that high Class Homophily Variance (CHV) explains why vanilla GNNs perform poorly in graph anomaly detection, since benign nodes are highly homophilic while anomalies are not. They propose HEAug, which samples homophily adjacency matrices from scratch with a self-attention mechanism, connects feature-similar but unlinked nodes, and modifies the loss to penalize unnecessary heterophilic edges. HEAug achieved the best results across eight benchmark datasets.

The paper proposes WTAC, weighted threshold anonymous credentials with redactable fine-grained blind signatures, for decentralized blockchain systems. It combines unlinkable redactable signatures, functional encryption, and weighted ramp secret-sharing so that issuers stay hidden from users and verifiers. The authors demonstrate the scheme in a privacy-preserving auditable lending system, give formal security proofs, and report performance evaluation showing practicality and efficiency.

Researchers propose MAD-ODE, a multivariate time series anomaly detection method for IoT settings that uses a learnable spatio-temporal graph ordinary differential equation (STGODE) network. It combines a static similarity graph built from prior knowledge with a learnable graph optimized through backpropagation to model sensor dependencies. The authors report that MAD-ODE achieves the best average performance and stable results across five datasets compared with existing methods.

Researchers propose LibPass, presented as the first adversarial attack against Android third-party library (TPL) detection tools. It generates adversarial apps with perturbations beyond standard obfuscation, using an improved firefly algorithm to search for them. Across five state-of-the-art detection tools and three datasets, LibPass achieved an average attack success rate of 61.33%, with a peak of 99.46%.

Researchers propose a behavioral analysis framework for detecting advanced web bots, specifically those built on OpenWPM, which can conceal browser fingerprints and imitate human browsing. The approach analyzes mouse movement, mouse clicks, keystrokes and scrolling using an ensemble of lightweight classifiers, with features augmented by unsupervised clustering. The method reportedly achieves an F1-score of 98.8%.

The paper proposes CTCV, a Collusion-resistant and Time-aware Collaborative Verification framework for auditing Edge Data Integrity in Mobile Edge Computing. It uses blockchain so edge nodes can verify data without trust dependencies, and a time-constrained challenge-response mechanism that sets a time threshold T between verification request and integrity proof inspection to accommodate node heterogeneity. The authors report that CTCV is computationally and communicationally more efficient than three baselines: EdgeWatch, EDI-S and EDI-V.

The authors propose QC-EAM, a quantum encryption and Quantum Fourier Transformation (QFT) based model for secure data communication and aggregation of smart grid power consumption data. The scheme encrypts user data with truly random keys and uses group homomorphism of Clifford Gates to evaluate encrypted quantum states. It was evaluated through classical simulations and on an actual quantum system provisioned through the IBM Qiskit platform, with the security analysis claiming perfect security via maximally mixed states.

The paper introduces MORAL, a model-based offline reinforcement learning method that replaces fixed-horizon rollouts with adversarial data augmentation. An adversarial process selects ensemble models against the policy for biased sampling, and a differential factor regularizes extrapolation error. The authors report that MORAL outperforms other model-based offline RL methods on the D4RL benchmark in policy learning and sample efficiency.

The paper proposes HIMT-NAS, an evolutionary multitask neural architecture search algorithm. It records each generation's operation and topology information to guide the search and prevent redundant exploration, and it adjusts cross-task knowledge transfer probabilities by measuring task similarity. Experiments on MedMNIST, CIFAR-10, CIFAR-100, and Tiny-ImageNet show consistent gains over single-task NAS and recent MT-NAS methods.

OpenAI launched an alignment research blog on December 1, 2025, as a lab notebook for sharing early-stage, narrow, or fast-moving work that is too preliminary for a full paper. The posts aim to be rigorous and technically sound, and the company wants feedback from the research community on ideas it is still exploring. The blog's focus is safety and alignment for increasingly capable AI, including systems capable of recursive self-improvement.

Researchers found that password strength meters in mainstream IoT systems overestimate passwords containing digit segments. They extracted high-frequency digit semantic patterns and used them to improve a PCFG attack, showing that these semantic characteristics strongly affect password strength. The authors also propose an improved password strength meter for IoT systems based on those patterns.