Loading vulnerability…
CVE-2026-12484 is an unsafe deserialization vulnerability in keras-team/keras version 3.15.0 affecting the public method keras.layers.TorchModuleWrapper.from_config. The method can deserialize attacker-controlled PyTorch pickle data by invoking torch.load(..., weights_only=False) without requiring an explicit unsafe opt-in. When the method is executed outside a SafeModeScope(True) context, no ambient safe-mode protection is enforced, so deserialization proceeds in an unsafe manner by default. Because Python pickle-based deserialization can execute attacker-controlled code during object reconstruction, processing an untrusted Keras layer configuration through this code path can result in 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.
keras.layers.TorchModuleWrapper.from_config. Where use of this functionality is unavoidable, ensure execution occurs within SafeModeScope(True) or an equivalent guard that enforces safe deserialization behavior. Restrict model and configuration inputs to trusted sources, isolate model-loading operations in hardened environments, and avoid deserializing attacker-controlled PyTorch pickle content.Patch, then assume compromise.
TorchModuleWrapper.from_config, including preventing implicit use of torch.load with weights_only=False unless an explicit unsafe opt-in is provided. Review application code and model-loading workflows to ensure untrusted serialized content is never passed through unsafe deserialization paths.No public exploits tracked yet. Mallory keeps watching.
No public exploit code observed for this vulnerability.
11 sources tracked across advisories and community write-ups. News coverage will land here when it surfaces.
No news coverage yet. Advisories and community discussion only.
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.