AI coding assistants' "GhostApproval" flaw allows unauthorized file access

A security flaw named "GhostApproval" allows AI coding assistants to access files outside their designated workspace, potentially leading to remote code execution. The vulnerability exploits symbolic links, tricking AI agents into writing to sensitive files like SSH keys. Some vendors fixed the issue, while others, including Anthropic, deemed it outside their threat model. This highlights a debate on whether AI tools should protect users from deceptive workspaces or if recognizing such threats is solely the user's responsibility.