New Jailbreaks Allow Users to Manipulate GitHub Copilot

Researchers demonstrated two practical weaknesses in GitHub Copilot: embedding conversational prompts inside code to trick the assistant into producing malicious code, and overriding Copilot's proxy settings to intercept authentication tokens and system prompts—potentially enabling unrestricted access to underlying LLMs, modification of safety constraints, and exposure of prompt histories. GitHub disputes the classification of these as vulnerabilities, but the techniques pose risks for malware generation, privacy leaks, and abuse of model access.