2026-11-12 · Microsoft Security · MEDIUM

Poisoned AI Training Enables Remote Code Execution

60-WORD SECURITY BRIEF

Poisoned training data and model checkpoints can trigger remote code execution in AI coding agents, per Microsoft Defender Research. The flaw exploits the training pipeline, letting attackers embed malicious code that runs when the agent processes user inputs, potentially compromising downstream systems. Defenders should verify training data integrity, enforce supply‑chain controls, and monitor agent behavior for anomalous execution.