AI Sec Watch

Deep model fusion is a technique that combines parameters or predictions from multiple deep learning models into one unified system to improve performance by reducing individual model biases and errors. The survey categorizes four main fusion approaches: weight average (averaging model parameters), mode connectivity (connecting models through optimized paths), alignment (matching corresponding units between models), and ensemble learning (combining model outputs during inference). However, applying this technique to large-scale models like LLMs (large language models, which are AI systems trained on massive amounts of text) faces challenges including high computational cost and interference between different types of models.

This paper presents SUNG, a framework for offline-to-online reinforcement learning (RL), which is training an AI agent first on existing data and then improving it through live interactions. The framework addresses two main problems: limited exploration due to offline data constraints and distribution shift (when the agent encounters data patterns it wasn't trained on). SUNG uses uncertainty estimation via a VAE (variational autoencoder, a type of neural network that learns data patterns) to guide both exploration (trying new actions) and exploitation (using known good actions), achieving strong performance on standard benchmarks.

A WordPress plugin called 'The AI Engine for WordPress: ChatGPT, GPT Content Generator' has a vulnerability that allows attackers with Contributor-level access or higher to read any file on the server. The problem exists because the plugin doesn't properly check file paths that users provide to certain functions (the 'lqdai_update_post' AJAX endpoint and the insert_image() function), which could expose sensitive information.

Google's new Antigravity IDE inherits multiple security vulnerabilities from the Windsurf codebase it was licensed from, including remote command execution (RCE, where an attacker can run commands on a system they don't own) via indirect prompt injection (tricking an AI by hiding instructions in its input), hidden instruction execution, and data exfiltration. The IDE's default setting allows the AI to automatically execute terminal commands without human review, relying on the language model's judgment to determine if a command is safe, which researchers have successfully bypassed with working exploits.

This research studies how deep neuro-fuzzy systems (DNFS, a type of AI that combines deep learning with fuzzy logic, which handles uncertain or imprecise information) perform on medical images that contain noise (unwanted degradation that makes images unclear). The researchers tested the DNFS on seven different medical imaging datasets with six types of noise and adversarial attacks (deliberate perturbations designed to fool AI models), and found that the DNFS maintained better accuracy on noisy images compared to other state-of-the-art models, though both approaches remained vulnerable to adversarial attacks.

Deep neural networks can be fooled by adversarial attacks (small, carefully crafted changes to input data that cause incorrect predictions), but training them to resist these attacks usually requires large amounts of labeled data. This paper proposes margin-based interpolation, a technique that adjusts how strongly to attack training data based on each example's difficulty and reliability, and uses global epsilon scheduling (gradually increasing perturbation strength during training) to help models become robust while maintaining accuracy, even with limited labeled data.

Researchers developed OR-SLZNet, a deep learning model that helps drones automatically identify safe landing zones by analyzing camera images in real time. The model assigns each pixel a safety score by combining visual features like color and texture with geometric information like flatness and slope, enabling drones to make quick landing decisions in emergencies or autonomous missions.

LangChain, a framework for building AI agents and applications powered by large language models, has a template injection vulnerability (a security flaw where attackers can hide malicious code in text templates) in versions 0.3.79 and earlier and 1.0.0 through 1.0.6. Attackers can exploit this by crafting malicious template strings that access internal Python object data in ChatPromptTemplate and similar classes, particularly when an application accepts untrusted template input.

Roo Code is an AI-powered coding agent that runs inside code editors. Before version 3.26.7, a validation error allowed Roo to automatically execute commands that weren't on an allow list (a list of approved commands), which is a type of command injection vulnerability (where attackers trick a system into running unintended commands).