AI Sec Watch

TensorFlow (an open-source machine learning framework) has a vulnerability in its `OpLevelCostEstimator::CalculateOutputSize` function where an integer overflow (when a calculation produces a number too large for the system to handle) can occur if an attacker creates an operation with tensors (multi-dimensional arrays of numbers) containing enough elements. The vulnerability can be triggered either by using many dimensions or by making individual dimensions large enough to cause the overflow.

TensorFlow, an open-source machine learning framework, has a vulnerability in its `OpLevelCostEstimator::CalculateTensorSize` function that can be exploited through integer overflow (a type of bug where numbers become too large for the program to handle correctly). An attacker could trigger this by creating an operation with a tensor (a multi-dimensional array of data) containing an extremely large number of elements.

TensorFlow, an open-source machine learning framework, has a typo in its `SpecializeType` code that causes a heap OOB (out-of-bounds, where the program tries to read or write memory outside the area it's allowed to access) read/write vulnerability. Due to the typo, a variable called `arg` uses the wrong loop index, which allows code to read and modify data outside the intended memory bounds.

TensorFlow's `AssignOp` (a copy operation in machine learning code) has a bug where it can copy uninitialized data (memory with random or leftover values) to a new tensor, causing unpredictable behavior. The code only checks that the destination is ready, but not the source, leaving room for uninitialized data to be used.

TensorFlow (an open source machine learning framework) has a bug where it sometimes fails to determine data types correctly during shape inference (the process of figuring out what dimensions data will have). The bug is hidden in production builds because assertion checks are disabled, causing the program to crash when it tries to use an error result as if it were valid data.

TensorFlow (an open source machine learning framework) has a vulnerability where attackers can crash TensorFlow processes by sending specially crafted data with invalid tensor types or shapes during decoding from protobuf (a data format used to serialize structured data). This is a denial of service attack, meaning the attacker can make the system stop working rather than gain unauthorized access.

TensorFlow, an open-source machine learning framework, has a bug where it can crash or behave unpredictably when decoding certain data structures (protobuf, a format for storing structured data) if some required information is missing. The problem occurs because the code only checks for this issue in debug builds (test versions), not in production builds (versions used in real applications), so real users may experience crashes or undefined behavior.

TensorFlow, an open-source machine learning framework, has a vulnerability in its Grappler component where the `set_output` function can write data to an array at any index specified by an attacker, creating a heap OOB write (out-of-bounds write, where data is written to memory locations it shouldn't access). This gives a malicious user the ability to write arbitrary data to unintended memory locations.

TensorFlow (an open-source machine learning framework) has a vulnerability where an attacker can crash the system by modifying a SavedModel file on disk to contain duplicate operation attributes, triggering an assertion failure (a built-in check that causes the program to stop if a condition is false). This is a denial of service attack (making a system unavailable to legitimate users).