Loading vulnerability…
A vulnerability in the huggingface/transformers library allows arbitrary code execution via deserialization of untrusted data in the load_repo_checkpoint() function of the TFPreTrainedModel class. The function uses pickle.load() on data that may originate from untrusted sources, enabling attackers to craft malicious serialized payloads. When such a payload is loaded, arbitrary code or commands can be executed on the victim's machine, typically during the model checkpoint loading process.
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 weaponized exploit for CVE-2024-3568, a remote code execution vulnerability in Hugging Face Transformers (<=4.37.2) when loading a malicious checkpoint via the TFAutoModel.load_repo_checkpoint() method. The exploit works by generating a checkpoint directory containing a crafted extra_data.pickle file with a pickle payload that executes arbitrary OS commands upon deserialization. The main.py script automates the creation of such malicious checkpoints, allowing the attacker to specify the target model and command to execute. The poc.py and src/payload/CVE_2024_3568/poc.py files demonstrate how loading the malicious checkpoint triggers code execution. The exploit is operational and can be used to prove code execution (e.g., by opening the calculator app or creating a file) without causing real harm. The repository is structured with clear separation between payload generation, utility functions, and demonstration scripts. No network endpoints are involved; the attack vector is local, requiring the victim to load a malicious checkpoint file.
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.
No public activity tracked yet. Mallory keeps watching.
No public activity observed for this vulnerability.
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.