AI & LLM Vulnerabilities

TensorFlow, an open source machine learning framework, has a bug in its shape inference for the `ReverseSequence` operation where it doesn't properly check if the `batch_dim` parameter is a negative number, allowing it to read memory outside the intended array bounds (a heap OOB read, or out-of-bounds read that accesses invalid memory). While the code checks that `batch_dim` isn't larger than the input rank, it fails to reject negative values that are too extreme, which can cause the program to access memory before the start of the array.

TensorFlow, an open source machine learning framework, has a vulnerability in its shape inference for the `Dequantize` operation where the `axis` argument is not properly validated. An attacker can provide an unexpectedly large `axis` value that causes an integer overflow (when a number becomes too large and wraps around to a negative or incorrect value) when the code adds 1 to it.

TensorFlow, an open-source machine learning framework, has a bug in its `Dequantize` function where the `axis` parameter (which specifies which dimension to operate on) isn't properly validated. This allows attackers to read past the end of an array in memory, potentially causing crashes or exposing sensitive data through a heap OOB (out-of-bounds) access, which means reading memory locations outside the intended storage area.

pytorch-lightning (a popular machine learning library) contains a vulnerability related to deserialization of untrusted data (CWE-502, where a program unsafely processes data from an untrusted source, potentially allowing an attacker to run malicious code). The vulnerability was identified and reported through the huntr.dev bug bounty program.

Sockeye, an open-source tool for Neural Machine Translation (a type of AI that translates text between languages), had a security flaw in versions before 2.3.24 where it used unsafe YAML loading (a method to read configuration files without proper safety checks). An attacker could hide malicious code in a model's configuration file, and if a user downloaded and ran that model, the hidden code would execute on their computer.

Aim is an open-source tool for tracking machine learning experiments. Versions before 3.1.0 have a path traversal vulnerability (a type of attack where special sequences like '../' are used to access files outside the intended directory), which could allow attackers to read sensitive files like source code, configuration files, or system files on the server.

TensorFlow's `saved_model_cli` tool (a command-line utility for working with machine learning models) has a code injection vulnerability because it runs `eval` on user-supplied strings, which could allow attackers to execute arbitrary code on the system. The risk is limited since the tool is only run manually by users, not automatically.

TensorFlow (an open source machine learning platform) has a vulnerability where shape inference code for certain operations can be tricked into accessing invalid memory through a heap buffer overflow (where a program writes data beyond the allocated memory space). This happens because the code doesn't verify that certain input parameters have the correct structure before using them.

TensorFlow, an open source platform for machine learning, had a memory leak and use-after-free bug (a mistake where the program tries to access data after it has already been deleted) in its `CollectiveReduceV2` function due to improper handling of asynchronous operations. The vulnerability was caused by objects being moved from memory while still being accessed elsewhere in the code.

TensorFlow's boosted trees code (a machine learning feature for building multiple decision trees together) lacks proper input validation, allowing attackers to crash the system (denial of service, where a service becomes unavailable), read sensitive data from memory, or write malicious data to memory buffers. The TensorFlow developers recommend stopping use of these APIs since the boosted trees code is no longer actively maintained.