Trained to betray: How post-training supply chain attacks hijack coding agents (on another site)

Supply chain attacks are a real risk for everyone these days, and that applies to AI agents as well. Agents are easily convinced to download scripts and modules to make performing tasks easier, and in fact will often download NPM or PyPI packages with no prompting by or awareness on the part of the user.

Microsoft Defender Research spent time discovering how poisoned datasets could be used to backdoor AI agents, and how those poisoned datasets could be delivered via trusted supply-chain channels.

In your organization, planning how and what to give AI access to is critical. Agents can and will use whatever tools seem appropriate to do the job unless effective guardrails are carefully and deliberately put in place.

Full article here: https://www.microsoft.com/en-us/security/blog/2026/11/12/trained-to-betray-how-post-training-supply-chain-attacks-hijack-coding-agents/