Blog
Published
An agent with shell access will eventually try a command you wish it hadn't. Use layers.
Layer 1: the server #
Turn on branch protection on GitHub or GitLab. Disable force pushes on main and release branches and require pull requests. This holds even if an agent tries.
Layer 2: each agent's own settings #
- **Claude Code:** deny rules in settings and PreToolUse hooks that inspect a command before it runs ([Claude Code hooks](https://gethrbr.com/blog/claude-code-hooks) ).
- **Codex:** execution rules that forbid a command or require approval ([a forbidden Codex rule](https://gethrbr.com/blog/why-codex-ignores-agents-md) ).
- Cursor: use the Allowlist run mode, so only the commands you list run without approval. Cursor's docs call the allowlist best-effort, not a security boundary. The Cursor CLI also takes deny rules.
The catch: these are set per tool and per machine. With a mixed team, the lists drift.
Layer 3: containment #
Run agents in a container, a disposable VM or a fresh clone. Keep backups. Containment limits damage when a rule is missed.
Where Harbor fits #
Harbor holds an agent to your rules at the moment it runs a command, and tells the agent what to do instead.
Example: your team bans prisma migrate reset because it wipes every row. You turn on a guardrail for it and write the remedy on it: a migration made with prisma migrate dev. Harbor blocks the command and points the agent to the remedy. The banned command never runs.
That last part matters. A bare “denied” leaves the agent guessing, and it often retries a variant. A block that names the alternative gets the task done.
You also do not start the list from a blank page. Harbor suggests guardrails from what your team already wrote: a rule that names the command, a deny entry in a repo's .claude/settings.json, or a package script that runs it. A person turns each one on, and it reaches Claude Code, Codex, Cursor and GitHub Copilot on every teammate's machine (guardrails).
Caveats #
- Harbor does not replace server-side branch protection or sandboxes. Keep both for the worst case. Harbor's hook fails open: if it errors, the call runs.
- Check your own setup by trying a banned command in a test repo before you rely on it.
harbor check "git push --force"prints which guardrail stops a command and what the agent will read.
Try it #
Run npx harborloop init. It is free for you and one teammate, no card.
Questions #
How do I stop an AI coding agent from running rm -rf or git push --force?
Use layers. Turn on branch protection so force pushes to main fail on the server. Set each agent's own controls: deny rules and PreToolUse hooks in Claude Code, forbidden execution rules in Codex, the Allowlist run mode in Cursor. Run agents in a container, a disposable VM or a fresh clone. Harbor adds one guardrail that refuses the command in Claude Code, Codex, Cursor and Copilot and tells the agent what to run instead.
Does Cursor have a deny list for terminal commands?
Not in the IDE, by its current docs. The run modes are Auto-review, Allowlist and Run Everything, and the docs call the allowlist best-effort, not a security boundary. The Cursor CLI takes deny rules in its permissions config.
Why tell the agent what to run instead of only denying the command?
A bare denial leaves the agent guessing, and it often retries a variant of the same command. A refusal that names the sanctioned route, such as a migration made with prisma migrate dev instead of prisma migrate reset, gets the task done.