Loading vulnerability…
CVE-2024-3660 is an unsafe deserialization vulnerability in TensorFlow-based Keras model handling prior to version 2.13. Loading a malicious model containing a Lambda layer can cause arbitrary code embedded in the model to be executed with the permissions of the application performing the load operation.
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.
This repository provides a proof-of-concept exploit for CVE-2024-3660, an arbitrary code execution vulnerability in TensorFlow Keras (< 2.13.0). The exploit consists of a Dockerfile that builds a malicious Keras .h5 model file containing a reverse shell payload. The payload is embedded via a Lambda layer in the model, which executes arbitrary shell commands when the model is loaded in a vulnerable environment. The README.md explains the vulnerability, affected versions, and usage instructions. The exploit requires the victim to load the malicious model file in a vulnerable TensorFlow environment, at which point a reverse shell is established to the attacker's specified LHOST and LPORT. The repository contains two files: a Dockerfile (which automates the creation of the malicious model) and a README.md (which documents the exploit and usage). The main fingerprintable endpoints are the generated model file, the temporary named pipe used for the shell, and the URL for downloading the TensorFlow wheel. The exploit is operational and demonstrates real-world impact by providing a working reverse shell payload.
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.
4 sources tracked across advisories, community write-ups, and news. New activity surfaces here as Mallory finds it.
A prior Keras vulnerability involving arbitrary code execution via Lambda layers.
A vulnerability referenced in the context of TensorFlow/Keras downgrade attacks (notably involving the TFLite converter) that can enable use of older, potentially unsafe model formats/behaviors leading to code execution risk.
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.