AI Sec Watch

TensorFlow, an open-source machine learning platform, has a vulnerability in TFLite (TensorFlow Lite, a lightweight version for mobile devices) where operations using quantization (a technique that reduces model size by using lower-precision numbers) can accidentally use uninitialized values because the code doesn't properly check whether quantization settings are valid before using them. This could cause unpredictable behavior in machine learning models running on mobile or embedded devices.

TensorFlow has a vulnerability where nesting `tf.map_fn` (a function that applies operations to tensor elements) calls with RaggedTensor inputs (tensors with variable row lengths) and no function signature can leak uninitialized memory from the heap and potentially cause data loss. The bug occurs because the code doesn't verify that inner tensor shapes match when converting from a Variant tensor to a RaggedTensor.

TensorFlow and Keras had a security flaw where loading machine learning models from YAML files (a text format for storing data) could let attackers run arbitrary code (any commands they want) on a system. The problem was caused by using an unsafe YAML parser that doesn't validate what code it runs.

TensorFlow, an open-source machine learning platform, has a vulnerability in its shape inference code for the `tf.raw_ops.Dequantize` function that could crash a system (denial of service via segfault, which is when a program crashes due to accessing invalid memory) if an attacker provides invalid arguments. The bug exists because the code doesn't properly validate the `axis` parameter before using it to access tensor dimensions (the size measurements of data structures in machine learning).

TensorFlow, an open-source machine learning platform, has a vulnerability where attackers can cause a denial of service (making a system unavailable by crashing it) through a segmentation fault (a memory error that crashes a program) in the MaxPoolGrad operation due to missing input validation on certain data structures called tensors. The vulnerability exists because an earlier fix for a related issue was incomplete.

TensorFlow, an open-source machine learning platform, has a vulnerability where attackers can crash the system (denial of service, a type of attack that makes a service unavailable) through a function called `tf.raw_ops.MapStage` because it doesn't validate that the `key` input is a proper non-empty tensor (a multi-dimensional array of numbers). This bug affects multiple versions of TensorFlow.

TensorFlow, an open source machine learning platform, has a vulnerability where an attacker can read data outside the intended memory bounds (a heap overflow, which is when a program accesses memory it shouldn't) by sending specially crafted invalid arguments to a function called tf.raw_ops.SdcaOptimizerV2. The vulnerability exists because the code doesn't verify that the length of input labels matches the number of examples being processed.

TensorFlow, an open source machine learning platform, has a vulnerability where attackers can read data outside the intended memory bounds by sending specially crafted arguments to certain functions like `tf.raw_ops.UpperBound` and `tf.raw_ops.LowerBound`. The vulnerability exists because the code doesn't properly validate the rank (the number of dimensions) of the input data it receives. This could allow attackers to access sensitive information stored in memory.

TensorFlow, an open-source machine learning platform, has a vulnerability in its `tf.raw_ops.NonMaxSuppressionV5` function that allows attackers to crash applications by supplying a negative number, which causes a division by zero error due to improper type conversion (converting a signed integer to an unsigned integer).