🚨 AI/ML SECURITY — A NEW FEDML VULNERABILITY CAN TURN A MODEL FILE INTO A REMOTE CODE-EXECUTION PATH CVE-2026-90614 Published: September 14, 2026 CyberSignal Priority: 🟠 HIGH A newly published vulnerability affects FedML up to version 0.9.6. The issue sits inside the framework's: MQTT + S3 communication backend specifically the: http://S3Storage.read_model() function. The weakness is: UNSAFE DESERIALIZATION. That matters because model artifacts are not always "just data." If an AI/ML platform loads attacker-controlled serialized objects, loading the model can potentially become code execution. ATTACKER-CONTROLLED S3 OBJECT ↓ FedML retrieves model ↓ unsafe deserialization ↓ attacker-controlled behavior ↓ potential remote code execution The vulnerability is tracked as: CVE-2026-90614 CWE-502 — Deserialization of Untrusted Data Public vulnerability information says remote exploitation is possible. The project had reportedly been informed through an issue report, but the public record says there had not yet been a response when it was published. 🛡️ Defender action If FedML is part of your ML infrastructure: → don't blindly trust remote model artifacts → restrict who can write to model-storage buckets → verify artifact integrity → isolate model-loading workers → minimize runtime privileges → monitor upstream remediation 🧠 CyberSignal insight The AI supply chain doesn't end at source code. A model checkpoint can itself become executable attack surface when the surrounding framework deserializes it unsafely. Source: CVE-2026-90614 · NVD references · FedML issue tracker September 14, 2026