Menu

Your coding agent's worktree is not its security boundary
πŸ“°

Your coding agent's worktree is not its security boundary

DEV Community: productivityΒ·heftyΒ·14 days ago
#M9fPlmhm
#devtoproductivity#devto#dev#account#access
Reading 0:00
15s threshold

An agent opens a fresh worktree, completes a task, and hands back a diff. The files are separate. The result is easy to review. It's tempting to call that containment, but the worktree only separates where the task's edits go . It does not decide which other files the process can read, which credentials it can use, or where it can connect.

DeepSeek's dsh-superagent-prototype makes the distinction unusually explicit. Its README describes a research prototype that handles permitted GitHub requests (and optionally Feishu), gives work a separate session and directory, then returns results to the original thread. Its deployment guide describes the other half of that setup: the published presets use danger-full-access under the operator's OS account.…

Continue reading β€” create a free account

Join HashtagPLUS to read full articles, follow hashtags, vote, and join the conversation.