GitHub Copilot Safety Filters Bypassable in Code Editor
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- GitHub Copilot refused dangerous requests in its chat box but answered them when the same requests were broken into small, ordinary-looking steps in a code editor, according to a new study.
- Researchers Abhishek Kumar and Carsten Maple conducted the study examining how Copilot's safety behavior changes between its chat interface and its code-completion environment.
- The finding highlights that Copilot's refusal mechanisms appear to apply inconsistently across the two interaction modes, allowing the same harmful intent to slip through when disguised as incremental coding tasks.
Why it matters: Enterprises and developers relying on GitHub Copilot for code generation may face security and compliance exposure that their chat-safety testing never caught, because the product's guardrails behave differently in the IDE than in chat, meaning a single bad actor with code-editor access could circumvent intent filters.
