AI Sec Watch

Pickle files (Python's serialization format for saving objects) can be backdoored because they execute code through opcodes (instructions that control a virtual machine). Attackers can inject malicious commands into pickle files using tools like fickling, and when someone loads the file, the hidden code runs without interrupting the program's normal function. This is especially dangerous in shared environments like Google Colab, where an infected pickle file could give attackers access to a user's connected Google Drive.

Streamlit, a Python framework for building data applications, has a directory traversal vulnerability (a type of attack where an attacker uses specially crafted file paths to access files they shouldn't be able to reach) in versions before 1.11.1. An attacker could trick the Streamlit server into reading and returning sensitive files from the server's file system, such as logs or other confidential information.

CVE-2020-25459 is a vulnerability in WeBank FATE (Federated AI Technology Enabler, a system for training machine learning models across multiple parties) versions 0.1 through 1.4.2 that allows attackers to read sensitive information during the training process. The issue exists in a function called sync_tree in the hetero_decision_tree_guest.py file, which means attackers could access private data while the model is being trained.

TensorFlow's `saved_model_cli` tool (a utility for working with saved machine learning models) had a code injection vulnerability in versions before 2.9.0, 2.8.1, 2.7.2, and 2.6.4, which could allow an attacker to open a reverse shell (a backdoor connection giving remote control of a system). The vulnerability existed because the tool used `eval` (a function that executes text as code) on user input for compatibility with older test cases, but since the tool requires manual operation, the practical risk was limited.

TensorFlow, an open source platform for machine learning, had a bug in two signal processing functions (`tf.compat.v1.signal.rfft2d` and `tf.compat.v1.signal.rfft3d`) where missing input validation (checking that data meets expected requirements before processing) could cause the software to crash under certain conditions. The bug was fixed in versions 2.9.0, 2.8.1, 2.7.2, and 2.6.4.

TensorFlow, an open source machine learning platform, had a bug in versions before 2.9.0, 2.8.1, 2.7.2, and 2.6.4 where certain converted models would crash when loaded. The problem occurred because the code assumed that quantization (a technique to compress model size by reducing numerical precision) would always use scaling factors smaller than 1, but sometimes the scale was larger, causing the program to stop unexpectedly.

TensorFlow, an open source platform for machine learning, has a vulnerability in the `tf.histogram_fixed_width` function where it crashes if the input data contains NaN (Not a Number, a special floating point value representing undefined results). The crash happens because the code tries to convert NaN to an integer without checking for it first, and this bug only affects the CPU version of TensorFlow.

TensorFlow version 2.8.0 had a bug in the `TensorKey` hash function (a function that converts data into a fixed-size code for quick lookups), where it incorrectly used `AllocatedBytes()` (an estimate of memory used by a tensor, including referenced data like strings) to access the actual tensor data bytes. This caused crashes because `AllocatedBytes()` doesn't represent the real contiguous memory buffer, and certain data types like `tstring` contain pointers rather than actual values.

TensorFlow, an open source machine learning platform, had a bug in versions before 2.9.0, 2.8.1, 2.7.2, and 2.6.4 where assertion macros (special code blocks that check if conditions are true) incorrectly compared different data types, specifically `size_t` and `int` values (two different ways to store whole numbers). This type confusion could cause assertions to trigger incorrectly due to how the computer converts between these different number types.