AI Sec Watch

BentoML, a Python library for serving AI models, had a vulnerability (before version 1.4.34) that allowed path traversal attacks (exploiting file path inputs to access files outside intended directories) through its configuration file. An attacker could trick a user into building a malicious configuration that would steal sensitive files like SSH keys or passwords and hide them in the compiled application, potentially exposing them when shared or deployed.

Adversarial imitation learning (AIL, a technique where an AI learns to mimic expert behavior by competing against a discriminator network) has worked well in practice but lacked solid theoretical foundations except in oversimplified settings. This paper introduces OPT-AIL (optimization-based adversarial imitation learning), a new framework that works with general function approximation (flexible neural network models rather than simple lookup tables), and proves it can learn expert-level policies efficiently while remaining practical to implement.

This research addresses multi-object navigation (MON), where an AI agent must find multiple targets in unknown environments by balancing immediate actions with long-term planning. Current methods focus too much on local path optimization, causing slow learning and getting stuck in trap states. The researchers propose GDWO (Goal-oriented Dynamic Weight Optimization), an algorithm that dynamically adjusts how much each target task contributes to the overall optimization by using gradient-based updates (mathematical techniques that improve decisions step-by-step) and normalizing weights based on navigation success rates, which improves learning efficiency and path planning.

The Kalrav AI Agent plugin for WordPress (versions up to 2.3.3) has a vulnerability in its file upload feature that fails to check what type of file is being uploaded. This allows attackers without user accounts to upload malicious files to the server, potentially leading to RCE (remote code execution, where an attacker can run commands on a system they don't own).

ChatterMate, a no-code AI chatbot framework (software that lets people build chatbots without writing code), has a security flaw in versions 1.0.8 and earlier where it accepts and runs malicious HTML/JavaScript code from user chat input. An attacker could send specially crafted code (like an iframe with a javascript: link) that executes in the user's browser and steals sensitive data such as localStorage tokens and cookies, which are used to keep users logged in.

This article argues that training AI models on copyrighted works should be protected as fair use (the legal right to use copyrighted material without permission for certain purposes like research or analysis), just as courts have previously allowed for search engines and other information technologies. The article contends that AI training is transformative because it extracts patterns from works rather than replacing them, and that expanding copyright restrictions on AI training could harm legitimate research practices in science and medicine.

This survey examines how rewards (scoring systems that guide AI behavior) are designed to align LLMs (large language models, or AI systems trained on massive amounts of text) with what humans want them to do. The review organizes the field by asking how rewards are mathematically defined, how they are built using different data sources and methods, how they work with different training approaches like reinforcement learning from human feedback (RLHF, a technique where humans rate AI outputs to improve performance), and how they are tested for safety and effectiveness.

Langflow contains a remote code execution (RCE, where an attacker can run commands on a system they don't own) vulnerability in its disk cache service that allows authenticated attackers to execute arbitrary code by sending maliciously crafted data that the system deserializes (converts from stored format back into usable objects) without proper validation. The flaw exploits insufficient checking of user-supplied input, letting attackers run code with the permissions of the service account.

Langflow, a workflow automation tool, has a vulnerability where attackers can inject malicious Python code into Python function components and execute it on the server (RCE, or remote code execution). The severity and how it can be exploited depend on how Langflow is configured.