CVE-2026-68770 is a security control bypass vulnerability in Hugging Face's sentence-transformers package affecting versions up to and including 5.5.1. The flaw is in the import_module_class helper in sentence_transformers/util/misc.py. A logic error in the trust gate includes a condition equivalent to accepting any model path that exists on the local filesystem, which causes the protection expected from trust_remote_code=False to be bypassed for local model directories. If an attacker can control or influence the contents of a model directory on disk, they can place malicious Python modules, including modeling-related files referenced through modules.json. When an application loads the model with SentenceTransformer(path, trust_remote_code=False), those attacker-supplied modules are imported and executed during the loading process. The issue therefore breaks the documented security boundary associated with disabling remote code trust and can lead to arbitrary code execution.
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.
No public exploits tracked yet. Mallory keeps watching.
No public exploit code observed for this vulnerability.
8 sources tracked across advisories, community write-ups, and news. New activity surfaces here as Mallory finds it.
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.