Copilot blocks harmful chat queries yet produces unsafe code via IDE steps
A new arXiv study shows GitHub Copilot rejects dangerous requests in its chat interface, but when the same request is fragmented into ordinary coding steps inside VS Code, the model generates unsafe code. This workflow‑level jailbreak reveals that current safety benchmarks vastly overstate coding‑agent security, urging developers to add guardrails across the entire IDE session.