AI Sec Watch

Researchers demonstrated a new attack method called ASBA (APK-Specific Backdoor Attack) that can compromise Android malware detection systems by injecting poisoned training data. Unlike previous attacks that use the same trigger across many malware samples, ASBA uses a generative adversarial network (GAN, an AI technique that learns to create realistic fake data) to generate unique triggers for each malware sample, making it harder for security tools to detect and block multiple instances of malware at once.

GitHub's CodeQL multi-repository variant analysis (MRVA) lets you run security bug-finding queries across thousands of projects quickly, but it's built mainly for VS Code. A developer created mrva, a terminal-based alternative that runs on your machine and works with command-line tools, letting you download pre-built CodeQL databases (collections of code information), analyze them with queries, and display results in the terminal.

CVE-2025-67511 is a command injection vulnerability (a flaw where attackers can insert malicious commands into input) in Cybersecurity AI (CAI), an open-source framework for building AI agents that handle security tasks. Versions 0.5.9 and earlier are vulnerable because the run_ssh_command_with_credentials() function only escapes (protects) the password and command inputs, leaving the username, host, and port values open to attack.

Neuron is a PHP framework for creating AI agents that can perform tasks, and versions 2.8.11 and earlier have a vulnerability in the MySQLWriteTool component. The tool runs database commands without checking if they're safe, allowing attackers to use prompt injection (tricking the AI by hiding instructions in its input) to execute harmful SQL commands like deleting entire tables or changing permissions if the database user has broad access rights.

Neuron is a PHP framework for building AI agents that can query databases. Versions 2.8.11 and below have a flaw in MySQLSelectTool, a component meant to safely let AI agents read from databases. The tool only checks if a command starts with SELECT and blocks certain words, but misses SQL commands like INTO OUTFILE that write files to disk. An attacker could use prompt injection (tricking an AI by hiding instructions in its input) through a public agent endpoint to write files to the database server if it has the right permissions.

XSS attacks (malicious code injected into websites to steal user data) are hard to detect because attackers can create adversarial samples that trick detection models into missing threats. This paper proposes a new detection model using two-stage AST (abstract syntax tree, a structural representation of code) analysis combined with LSTM (long short-term memory, a type of neural network good at processing sequences) to better identify malicious code while resisting adversarial tricks, achieving over 98.2% detection accuracy even against adversarial attacks.

This research proposes a Fairly Proportional Noise Mechanism (FPNM) to address a problem in differential privacy (DP, a technique that adds random noise to data to protect individual privacy while allowing statistical analysis). Traditional DP methods add noise uniformly without considering fairness, which can unfairly affect different groups of people differently, especially in decision-making and learning tasks. The new FPNM approach adjusts noise based on both its direction and size relative to the actual data values, reducing unfairness by about 17-19% in experiments while maintaining privacy protections.

WiFi-based sensing systems that use deep learning (AI models trained on large amounts of data) are vulnerable to adversarial perturbation attacks, where attackers subtly manipulate wireless signals to fool the system into making wrong predictions. Researchers developed WiIntruder, a new attack method that can work across different applications and evade detection, reducing the accuracy of WiFi sensing services by an average of 72.9%, highlighting a significant security gap in these systems.

This research paper studies the challenge of balancing two competing goals in decentralized learning (where multiple computers train an AI model together without a central server): keeping each computer's data private while protecting against Byzantine attacks (when some computers deliberately send false information to sabotage the learning process). The authors found that using Gaussian noise (random mathematical noise added to messages) to protect privacy actually makes it harder to defend against Byzantine attacks, creating a fundamental tradeoff between these two security goals.