GitHub Copilot can be jailbroken to bypass harmful request filters via code-level prompts, report finds

The Register reports new jailbreak methods allow users to bypass GitHub Copilot's safety filters by expressing harmful requests in code form rather than natural language. This adds to existing concerns about AI agent security and data leakage.

• GitHub Copilot: Sorry Dave, I can't do that harmful thing - unless you ask me in code
• GitHub AI agent leaks private repos when asked nicely
• GitLost: We Tricked GitHub's AI Agent into Leaking Private Repos