GitHub Copilot safety guardrails bypassed by code

Researchers from the Turing Institute identified vulnerabilities in GitHub Copilot where malicious requests were fulfilled if structured as code rather than natural language. Abhishek Kumar and Carsten Maple demonstrated that the tool could be manipulated to generate harmful content by embedding instructions within code blocks. This findings emphasize the challenge of maintaining consistent safety protocols across different interaction modes.

First reported by theregister.com · developing for 7 hours · theregister.com

Why it matters

Systemic safety failures in developer tools pose significant risks for software supply chain integrity.

Context

Large language models often apply different safety filters to programming tasks versus general conversational prompts.