Harness details symlink timing bypass in AI permission checks
The author argues that AI coding assistants and permission systems often suffer from representation and timing gaps, validating permissions against path strings while the underlying filesystem actions resolve through symlinks or change before execution. This mismatch allows security bypasses, as demonstrated by vulnerabilities found in tools like Claude Code and Cursor. To achieve secure boundaries, permission checks must evaluate the actual resolved targets and files used by the system actions rather than relying solely on text paths. Open article →