Prevent destructive AI agent actions
AI coding agents have wiped production databases, deleted home directories, and force-pushed away weeks of work. Agent Approve adds policy checks and human review before supported actions run.
Some commands have no undo
AI agents move fast. Most of the time that is what you want. But when something goes wrong, it goes wrong in seconds — rm -rf, DROP DATABASE, git push --force — and the damage is often unrecoverable.
The worst incidents share three patterns:
- Admin-level commands deserve deliberation. A human operator would slow down before running
volumeDeleteon a production environment. A Cursor agent fired one against the PocketOS Railway environment in nine seconds. - Agents do not always follow instructions. The Replit agent that wiped a production database during a SaaStr demo had been told not to change code.
- A single hidden failure is permanent. A swallowed
rm -rf ... 2>/dev/nullcost one Claude Code user 1.5 weeks of work before they noticed.
Hooks let Agent Approve apply your policy before supported actions run. Lists decide known matches, Policy Rules evaluate other actions, and your default determines what still needs review on your iPhone or Apple Watch.
See destructive-action protection in action
Watch Agent Approve intercept a dangerous command and send the decision to a trusted device before anything runs.
Real incidents from real teams
These risks have real impact — from solopreneurs shipping a side project to teams running production for large companies. Together, the incidents below make the scope of the problem clear.
The problem is not going away as models improve. In July 2026, within days of the GPT-5.6 launch, developers reported a cluster of destructive-cleanup incidents — a wiped Mac home directory, five validators corrupted by an rsync --delete-delay against /, and files removed by a PowerShell -Include filter that matched more than intended. OpenAI’s own GPT-5.6 model card reports a lower data-overwrite-avoidance score for GPT-5.6 Sol than for GPT-5.5 (0.83 versus 0.88). Guardrails outside the model matter regardless of which model you run.
How Agent Approve protects you
Agent Approve sits between your AI agent and the rest of your computer. Every command runs through your policy first, and anything that needs human judgment goes to your phone or watch instead of the agent’s terminal session.
Hooks intercept commands before they execute
Agent Approve installs hook scripts or plugins for Cursor, Claude Code, Windsurf, Antigravity, Antigravity CLI, legacy Gemini CLI, GitHub Copilot CLI, OpenAI Codex, OpenHands, OpenCode, OpenClaw, Hermes, and more. Supported pre-action hooks let the service check requests before execution; coverage depends on the integration.
Your policy decides first
Restrictive Ask asks you about risky-command matches and allows other work by default, with a small Deny List for explicit catastrophic commands. Choose Restrictive Block to block the broader destructive-command list instead. You can enable AI Policy Rules separately.
Approval happens on a different device
Manual approvals show up on your iPhone or Apple Watch. The agent cannot click an "allow" button on your behalf because the answer happens on a separate trust path.
Push notifications, no laptop required
Walk the dog, grab coffee, take a meeting. When an approval is needed you get a push notification with the parsed command and one-tap approve or deny.
Every decision is logged
Event history shows approvals, denials, and automatic decisions. When Policy Rules run, event details include rule matches and the evaluator used, subject to your history and privacy settings.
One policy across every agent
Your active policy shares lists, enabled Policy Rules, default behavior, and approval timeouts across connected integrations in Approval Mode. You do not have to recreate the policy for each agent.
Policy Rules add another layer
A destructive action may arrive as a shell command, a database query, or an MCP tool with structured arguments. Policy Rules let you describe the condition you want checked, such as permanently deleting data or changing a production system, without listing every possible command.
Add rules from the library or write your own definition with matching and non-matching examples. Choose Deny, Ask, or Allow and set a match threshold. The evaluator returns a probability; Agent Approve applies your chosen action when the rule matches. Model evaluation can miss an action or raise a false alarm, so review results and tune the rule.
Lists run first. Policy Rules check actions the lists have not decided, then the policy default applies. A broad Allow List entry can bypass the rule you intended to rely on. Keep those entries narrow and review the whole policy together.
In the version 0.9 preview, Solo and Pro include Agent Approve Fast, powered by Jev and using AI credits, with SubQ Snap and Perplexity Decider through OpenRouter as included fallbacks. You can also use your own provider account. See Policy Rules for setup and data sharing, or compare plans and usage.
Command protection that ships with Agent Approve
The Restrictive Block policy ships with the patterns below already populated. You can edit, extend, or replace them from the iOS app. Every category traces back to a publicly reported incident or a known dangerous default.
List matching uses literal exact, prefix, and contains patterns. For supported shell tools, the command parser recognizes separators such as &&, ||, |, ;, and newlines, while respecting quoted text and backslash line continuations. A denied command in a parsed sequence blocks the whole request. Policy Rules provide a separate check for actions not decided by the lists. See the full decision order.
Deny evaluation also sees through privilege-escalation wrappers. A pattern that blocks shred also blocks sudo shred, sudo -u root shred, and doas or run0 variants of the same command. It works in the other direction too: a deny pattern you write as sudo rm also catches flagged forms like sudo -E rm and equivalent wrappers, without ever denying an unprivileged rm. The allow list deliberately does not work this way: approving git status never auto-approves sudo git status. An unmatched escalated command proceeds to enabled Policy Rules, then the policy default.
How to enable the Restrictive Block policy
Once Agent Approve is connected to your agents, switching to the Restrictive Block policy takes three taps in the iOS app.
- 1.Open the Agent Approve iOS app and tap Policy.
- 2.Tap the policy selector at the top of the screen.
- 3.Choose Restrictive Block. Its current lists apply to subsequent requests from connected integrations in Approval Mode. Review those lists if you have edited the built-in policy.
Known limitations
Lists match known patterns in the action reported by the integration. Policy Rules evaluate the supplied tool input and context. Neither can guarantee detection of destructive behavior or predict everything a command will do at runtime.
Start by reviewing the Restrictive Block policy and adding Policy Rules for conditions that matter to your work. Keep its Ask User default for undecided actions, and choose Ask me when rules cannot be checked. Review each request on your iPhone or Apple Watch before approving it.
- Pattern matching cannot catch every alias, wrapper, encoded payload, unusual flag order, or generated script.
- Some MCP and native tools express destructive actions as structured parameters instead of familiar shell commands.
- Local risk depends on your project, credentials, cloud accounts, mounted drives, and production access.
- You remain responsible for every command that runs on your machine, including ones you approve through Agent Approve. Please read our Terms of Service before relying on the Restrictive Block policy in production.
Review your lists and Policy Rules as your tools and projects change. Test changes with harmless actions, inspect the resulting decisions, and keep manual review for work you are not ready to approve automatically.
Set up destructive-action protection
Follow the task guide to activate Restrictive Block, verify the approval path without risking data, and add safeguards for your environment.
Open the setup guide