Loading vulnerability…
CVE-2023-43654 is a critical vulnerability in TorchServe affecting versions 0.1.0 through 0.8.1. TorchServe's default configuration does not properly validate user-supplied model source locations, allowing a user to instruct the service to fetch a model archive from an arbitrary remote URL and write attacker-controlled content to disk. Because TorchServe is designed to load and serve models, this behavior can be abused to introduce a malicious model into the serving environment. In practical exploitation scenarios, the weakness can be chained into remote code execution when the malicious model is loaded or otherwise processed by the service. The issue stems from insecure default handling of externally supplied URLs and insufficient restriction of permitted download locations unless administrators explicitly configure allowed URL controls.
Mallory correlates every CVE against your assets, your vendors, and active adversary campaigns. Know which vulnerabilities matter for you, not just which ones are loud.
What it means. What to do now. Patch path, mitigations, and the assume-compromise checklist.
What an attacker gets, and what they’ve been doing with it.
If you can’t patch tonight, do this now.
Patch, then assume compromise.
1 valid exploit after Mallory filtered fakes, detection scripts, and README-only repos (2 hidden).
This repository contains a single Metasploit module targeting PyTorch TorchServe servers vulnerable to a chain of issues culminating in unauthenticated remote code execution (RCE). The exploit leverages CVE-2023-43654 (model registration SSRF) and CVE-2022-1471 (SnakeYAML deserialization RCE). The module works by registering a model from an attacker-controlled server, which serves a malicious MAR file containing a crafted YAML configuration. This configuration exploits the deserialization vulnerability to load and execute arbitrary Java code on the target. The module includes checks for the TorchServe API and version, and cleans up by unregistering the model after exploitation. The main attack vector is network-based, targeting the HTTP management interface of TorchServe, typically on port 8081. The code is written in Ruby and is structured as a standard Metasploit exploit module.
Products and vendors Mallory has correlated with this vulnerability. Open in Mallory to drill down to specific CPE configurations and version ranges.
Vendor-confirmed product mapping. Mallory continuously reconciles this list against your asset inventory.
6 sources tracked across advisories, community write-ups, and news. New activity surfaces here as Mallory finds it.
A high-severity remote code execution vulnerability (part of a chain) in TorchServe (PyTorch model serving) enabling unauthorized access, malicious model insertion, and potential full server takeover.
A critical remote code execution vulnerability in PyTorch TorchServe (part of the ShellTorch set) that can allow complete unauthorized access; the content notes it can be reached/triggered via the 0.0.0.0 Day localhost-reachability issue in internal/local deployments.
Query your assets running an affected version, and investigate the blast radius.
Every observed campaign linking this CVE to a named adversary.
Malware families riding this exploit, with evidence and IOCs.
YARA, Sigma, Snort, and vendor rules, auto-deployed to your SIEM.
Cross-references every affected SKU, including bundled OEM variants.
Community discussion across Reddit, Mastodon, and other social sources.