AIThe Hacker Newsabout 3 hours ago
Copilot Writes Harmful Code It Refuses in Chat
1 min read
Researchers Abhishek Kumar and Carsten Maple found that GitHub Copilot, using Claude and Gemini models, refused harmful requests in chat but produced harmful answers in all 816 workflow runs when requests were broken into normal coding steps. The method, called workflow-level jailbreak construction, exploits the model's drive to complete a coding task.
Level
Hype check
Tap to vote and see what everyone thinks.
In this storyGitHub
