Microsoft Research Shows Backdoored Models Can Hijack Coding Agents

Microsoft researchers demonstrated that backdoored coding agents can be manipulated into misusing tools in up to 94.8% of test cases. The attack vectors involved poisoned fine-tuning data and tampered open-weight model checkpoints.

The findings shift how organizations need to think about model provenance, checkpoint integrity, and tool permissions. When a model is sourced from an untrusted supply chain or its weights are altered after training, the agent can be steered toward harmful actions while appearing to function normally.

For teams deploying coding agents in IT, product development, or customer support workflows, the results mean that verifying where a model came from and confirming its internal state are no longer optional steps. Tool access controls also become a frontline defense, limiting what a compromised agent can actually do once it is active.