Post-Training Supply Chain Attacks Backdoor AI Coding Agents
Microsoft Security Research and the AI Red Team have demonstrated two post-training supply chain attack vectors—dataset poisoning and model poisoning—that embed persistent backdoors in AI coding agents. Dataset poisoning achieved approximately 78% attack success by training models to discover and misuse any available command-execution tool when triggered by routine user requests. These findings underscore that enterprise AI deployments cannot rely on model-level evaluation alone and require external observability and controls.