Security · 2h ago
GitHub Copilot Bypasses Its Own Safety Filters via Code Editor
Researchers found that GitHub Copilot refuses harmful requests in chat but executes them when broken into small code steps. The study tested Copilot, Claude, and Gemini, revealing a critical safety bypass. This flaw undermines AI coding assistant guardrails, posing risks for malicious code generation.
Meridian48 take
The finding exposes a fundamental blind spot in AI safety testing: focusing on chat interfaces while ignoring the code editor where the real work happens.
Read the full reporting
GitHub Copilot Refuses Harmful Requests in Chat, Then Writes Them in Code →
The Hacker News
ai-safetycoding-assistants