CVE-2024-31583

Vulnerability

Overview

A use-after-free vulnerability exists in PyTorch versions prior to v2.2.0, specifically located in the file torch/csrc/jit/mobile/interpreter.cpp [1][2]. This flaw occurs when the interpreter improperly handles memory for certain operations in the mobile JIT execution path, leading to a dangling pointer that can be dereferenced after the associated memory has been freed [4]. The root cause is a missing reference count adjustment or insufficient lifetime management for an object used during instruction dispatch.

Attack

Vector

An attacker can trigger this vulnerability by crafting a malicious PyTorch model file (e.g., .pt or .pth) that, when loaded and executed via the mobile interpreter, causes the use-after-free condition [2]. No authentication is required if the victim loads the attacker-supplied model; the attack surface includes any application using PyTorch's JIT mobile interpreter to run untrusted models. The vulnerability is reachable without special privileges, as the interpreter runs at the same user level as the consuming application.

Impact

The primary impact is denial of service (application crash) due to memory corruption [2]. Under specific conditions, an attacker may achieve arbitrary code execution by controlling the freed memory's content or layout, though this depends on heap state and mitigations like ASLR. The CVSS base score is 7.3 (High), with vector AV:N/AC:L/PR:N/UI:N/S:U/C:L/I:L/A:L, indicating potential impacts to confidentiality, integrity, and availability [2].

Mitigation

The vulnerability was fixed in PyTorch version 2.2.0 [1][3]. Users are strongly advised to upgrade to v2.2.0 or later. No known workarounds exist; loading models from untrusted sources should be avoided until the update is applied. The flaw is not currently listed in CISA's Known Exploited Vulnerabilities catalog.