Researcher finds GitHub's AI agent can be tricked into exposing private repos

Colin Eberhardt discovered a vulnerability in GitHub's Copilot agent that allows attackers to manipulate it into leaking access to private repositories through social engineering. The finding highlights a real risk in AI-assisted development tools: when agents have broad permissions, subtle prompts can override intended safeguards.