Skip to content
Anthropic Claude Code·Indexed by this site · Originally publishedSelectedAttention40

Claude Code 2.1.299 fixes issue where sandbox commands writing to --plugin-dir were loaded and executed outside the sandbox

AI summary

Claude Code 2.1.299 fixes a sandbox privilege escalation issue: shell commands inside the sandbox could previously write to the --plugin-dir directory, and code in that directory could be reloaded and run by Claude Code outside the sandbox. The same version also fixes an issue on macOS where sandboxed commands could read or write denied paths via symlinks using their real paths, and a problem where the sandbox's allowRead and allowWrite wildcard rules were no longer revalidated after their starting directory was replaced by a symlink. Enforcement of permission rules for user-installed mods, managed hooks, and organization MCP tool names was also strengthened, and an issue where MCP servers were connected before some policy checks completed has been fixed.

Why it was selected

The changelog lists the sandbox escape fix alongside symlink path validation, so teams using coding agents can check the isolation boundary of the plugin directory accordingly.

Read originalgithub.com