AI Sec Watch

Tencent HunyuanDiT (an AI image generation model) has a remote code execution vulnerability in its model_resume function that allows attackers to run arbitrary code if a user opens a malicious file or visits a malicious page. The flaw stems from improper validation of user input during deserialization (converting data from storage format back into usable objects), allowing attackers to execute code with root-level privileges.

This research addresses machine unlearning in neural IR (information retrieval, the technology that ranks search results), a process called neural machine unranking (NuMuR) that selectively removes data from AI systems for privacy compliance. The authors propose CoCoL (contrastive and consistent loss, a method with two complementary training objectives), which uses a contrastive loss to reduce relevance scores on forgotten data while preserving performance on shared data, plus a consistent loss to maintain accuracy on retained data, demonstrating effective data removal across multiple neural ranking models.

CVE-2025-63664 is a flaw in the GT Edge AI Platform (before version 2.0.10-dev) where incorrect access control in the /api/v1/conversations/*/messages API allows attackers without permission to view other users' message histories with AI agents. This is classified as improper access control (CWE-284, a category of security flaws where systems fail to properly restrict what users can access).

Cyberbullying on social media is a growing problem that harms people's mental health, and traditional methods to stop it are no longer effective. This study examines how artificial intelligence can help protect online communities from cyberbullying by exploring different AI technologies, their uses, and the challenges involved. The goal is to understand how AI might create safer online environments.

Generative AI (systems that create new text, images, or other content) is transforming many industries but raises ethical concerns like data privacy (protecting personal information), bias (unfair treatment of certain groups), transparency (being open about how the AI works), and accountability (responsibility for the AI's actions). Researchers propose a trust framework based on transparency, fairness, accountability, and privacy to help ensure generative AI is developed and used responsibly.

Large language models (LLMs, AI systems trained on huge amounts of text to generate human-like responses) can now mimic not just general human language but also unusual, individual-specific human behaviors. This ability could lead to LLMs being used more widely in research studies and potentially reduce the role of actual humans, which raises concerns about AI alignment (ensuring AI systems behave in ways humans intend and approve of) and how this technology affects society.

Pre-trained language models (PLMs, large AI systems trained on text data) can be vulnerable to backdoor attacks, where hidden triggers in input cause the model to produce manipulated output. This paper proposes an enhanced defense method called masking-differential prompting (MDP) that works with few-shot learning (training on very small datasets), using Jensen-Shannon divergence (a mathematical measure to compare probability distributions) instead of traditional methods and an automatic threshold-selection approach to better detect and block these attacks.

This article presents a new method for detecting insider threats (malicious activities by trusted employees within an organization) by combining two types of graph structures (explicit graphs based on predefined rules and implicit graphs derived from data patterns) with neural network models (GCNs and Bi-LSTM, which are deep learning architectures that process network relationships and temporal sequences). The framework was tested on two datasets and achieved very high accuracy rates, correctly identifying 100% of threats on one dataset while maintaining a low false positive rate (incorrectly flagging normal activity as suspicious).

This research addresses a problem in federated learning (a method where multiple computers train an AI model together without sharing raw data) combined with adversarial training (a technique that makes AI models resistant to intentionally tricky inputs). The authors found that simply combining these two approaches causes the model's accuracy to drop because adversarial training increases differences in the data across different computers, making the federated learning less effective. They propose SFAT (Slack Federated Adversarial Training), which uses a relaxation mechanism to adjust how the computers combine their learning results, reducing the harmful effects of data differences and improving overall performance.