AI Sec Watch

MLflow, a machine learning platform, had an insecure temporary file vulnerability (CWE-377, a weakness where temporary files are created without proper security protections) in versions before 1.23.1. This vulnerability could potentially allow attackers to access or modify sensitive data stored in temporary files.

TensorFlow (an open source machine learning framework) has a vulnerability where building an XLA compilation cache (a storage system that speeds up machine learning model compilation) with default settings causes a null pointer dereference (a crash that happens when code tries to use a memory location that doesn't exist). This occurs because the default configuration allows all devices, leaving a critical variable empty.

TensorFlow (an open-source machine learning framework) has a vulnerability in its TFG dialect, which is part of MLIR (a compiler framework for optimizing code). An attacker can modify the SavedModel format (the way trained models are saved to disk) to break assumptions the system makes, which can crash the Python interpreter or cause heap OOB (out-of-bounds memory access, where code reads or writes memory it shouldn't).

TensorFlow, an open-source machine learning framework, has a vulnerability in its `simplifyBroadcast` function (a part of the MLIR-TFRT infrastructure, which is the compiler and runtime system) that causes a segfault (a crash from accessing invalid memory) when given scalar shapes (data without dimensions), resulting in a denial of service (making the system unavailable). This affects only TensorFlow version 2.7.0.

TensorFlow (an open-source machine learning framework) has a vulnerability where type inference can read data outside the bounds of allocated memory (a heap out of bounds read). The bounds checking uses a DCHECK, which is disabled in production code, allowing an attacker to manipulate a variable so it accesses memory beyond what is available.

TensorFlow (an open-source machine learning framework) has a vulnerability where the GraphDef format (TensorFlow's way of representing computation graphs) can accept self-recursive functions even though it shouldn't, causing a stack overflow (a crash from too much memory use) when the model runs because the system gets stuck trying to resolve the same function repeatedly.

TensorFlow (an open source machine learning framework) has a vulnerability where a maliciously altered GraphDef (a representation of a machine learning model's computation graph) from a SavedModel can crash a TensorFlow process by forcing extraction of a value from a StatusOr (a data structure that holds either a valid result or an error state). The issue affects both TensorFlow 2.7 and 2.8 versions.

TensorFlow, a machine learning framework, has a vulnerability (CVE-2022-23589) in its Grappler component (a graph optimization tool) that can cause a null pointer dereference (crash from accessing invalid memory) when processing maliciously altered SavedModel files (serialized machine learning models). The bug occurs in two places during optimization operations and can be triggered by missing required nodes in the computation graph.

A malicious user can crash TensorFlow (an open source machine learning framework) by modifying a SavedModel (a pre-trained model file) in a way that tricks the Grappler optimizer (a tool that improves model performance) into building a tensor with an invalid reference dtype (data type), causing the program to fail.