AI Sec Watch

ONNX (a machine learning model format library) versions before 1.13.0 contain a directory traversal vulnerability (a security flaw where an attacker can access files outside the intended folder by using paths like '../../../etc/passwd'). An attacker could exploit the external_data field in tensor proto (data structure in ONNX models) to read sensitive files from anywhere on a system.

The EU AI Act requires technical standards to be written by European standardization organizations (CEN and CENELEC) that explain how companies can safely build high-risk AI systems. These standards follow a six-step approval process and, once published and approved by the European Commission, become 'harmonized and cited standards' that legally presume compliance with safety regulations if companies follow them. The drafting process is currently ongoing but behind schedule, with different standards at different completion stages.

TensorFlow, an open source platform for machine learning, has a bug in the MakeGrapplerFunctionItem function where providing inputs larger than or equal to the output sizes causes an out-of-bounds memory read (reading data from memory locations the program shouldn't access) or a crash. The issue has been patched and will be included in TensorFlow 2.11.0 as well as backported to earlier versions.

TensorFlow, an open source machine learning platform, has a bug in its MakeGrapplerFunctionItem function where providing input sizes that are greater than or equal to output sizes causes an out-of-bounds memory read (accessing memory locations outside the intended range) or a crash. This vulnerability affects how TensorFlow processes data when sizes are mismatched.

This post demonstrates that ChatGPT can be prompted to roleplay as a Microsoft SQL Server (a database management system) and respond with realistic database commands and results, including creating databases, tables, inserting data, and writing stored procedures (reusable blocks of SQL code). The author shows that ChatGPT can understand user intent well enough to execute complex database operations like UPSERTs (operations that update existing records or insert new ones if they don't exist), even when given incomplete information.

PyTorch versions before trunk/89695 have a vulnerability in the torch.jit.annotations.parse_type_line function that can allow arbitrary code execution (running attacker-controlled commands on a system) because it uses eval unsafely (eval is a function that executes code from text input without proper safety checks).

TensorFlow, an open source platform for machine learning, has a bug where converting character data to boolean values can cause crashes because the conversion is undefined unless the character is exactly 0 or 1. This issue affects the process of printing tensors (multi-dimensional arrays of data used in machine learning).

TensorFlow (an open source machine learning platform) has a vulnerability where invalid input to a specific function causes a segfault (a crash where the program tries to access memory it shouldn't). The bug occurs when `tf.raw_ops.CompositeTensorVariantToComponents` receives an `encoded` parameter that is not a valid `CompositeTensorVariant` tensor (a data structure for machine learning computations).

TensorFlow, an open-source machine learning platform, has a vulnerability where passing a `token` input that is not UTF-8 encoded (a character encoding standard) causes the `tf.raw_ops.PyFunc` function to crash with a CHECK fail (a safety check that stops execution when something is wrong). This is a type of improper input validation weakness, meaning the function doesn't properly check whether its input is in the correct format before processing it.