Can OpenClaw Delete Your Emails? How to Stop It and Require Approval

Published 2026-11-15 · Updated 2026-11-15 · By The Is My Agent Safe team

Short answer

Yes. An OpenClaw agent with a tool that deletes mail or files can use it, even after you told it in chat to ask first. OpenClaw's docs say long chats get summarized and older details may drop out. To stop a run, send /stop. To make it ask first, deny delete tools by name and set exec approvals to ask.

  • A chat instruction is not a setting. OpenClaw's compaction docs say older details may leave the model's context.
  • /stop aborts the current run. Stopping the openclaw gateway process stops everything.
  • Tool policy works on tool names, and deny always wins. Deny delete tools you don't need.
  • Exec approvals decide whether shell commands run, ask a human, or are denied.
  • Re-run openclaw security audit --deep after every change.

Paste your audit: run openclaw security audit --deep and paste the output into our free check for a red/amber/green result with the most important finding first (Limits of the check).

Who this is for

People who let an OpenClaw agent manage an inbox, a folder or a database, and want it to ask before it deletes anything.

Can an OpenClaw agent really delete my emails?

Yes, if it has a tool that can. In February 2026, a Meta AI alignment director posted on X that her OpenClaw agent bulk-deleted mail from her real inbox. She had asked it to suggest what to archive or delete and not to act until she said so (Business Insider, Feb 2026). She said she couldn't stop it from her phone and had to stop the processes on her Mac mini (Windows Central, Feb 2026).

Why did the agent forget "confirm before acting"?

She said the large inbox triggered compaction, and her instruction was lost (Windows Central). OpenClaw's compaction docs explain the mechanism. Long conversations are summarized to stay within model limits. A degraded summary is "deliberately lossy", and "older details, identifiers, and pending requests may therefore leave the model context."

So a rule you type in chat is not a setting. The steps below put the rule in OpenClaw's config, where the model's memory doesn't matter.

How do I stop a run that is deleting things?

  1. Send /stop in the chat. OpenClaw's slash-command docs say it aborts the current run and stops its ordinary Gateway or sandbox commands. Services started with background: true keep running, so stop those separately.
  2. If that doesn't work, stop it on the host. OpenClaw's incident response docs say to stop the macOS app if it supervises the Gateway, or terminate your openclaw gateway process.

Typing "stop" in plain words is just another message to the model. /stop is a command the Gateway handles.

How do I make OpenClaw ask before it deletes?

Step 1: See what the agent can do

On the Gateway host, run:

openclaw --version
openclaw exec-policy show --agent main
openclaw sandbox explain --agent main

Per OpenClaw's approvals CLI docs, exec-policy show explains the agent's terminal tool policies and command approvals, and never changes permissions.

Step 2: Deny the delete tools you don't need

OpenClaw's tool policy docs say deny always wins, and tool policy filters tools by name. MCP server tools carry a server prefix, such as the docs' example outlook__send_mail. Add the mail server's delete or trash tools to tools.deny for the agent that reads your inbox, then confirm they're gone in the agent's Tools settings.

Note that the messaging and coding profiles implicitly allow configured MCP servers, per the tool policy configuration docs. A restricted profile alone doesn't remove a mail server's delete tool.

Step 3: Make shell commands ask a human

Tool policy works on names. It does not inspect what a shell command does, so denying write doesn't make exec read-only. OpenClaw's exec docs list the modes: deny, allowlist, ask (allowlist matches run, everything else asks a human), auto and full (no ordinary prompts). For an inbox agent that doesn't need a shell, deny exec. Otherwise:

openclaw exec-policy set --security allowlist --ask on-miss

Those values come from the docs' list of accepted flags. Then run openclaw exec-policy show again to confirm.

Step 4: Use a stricter permission mode for the session

OpenClaw's permission modes include read-only (exec denied, mutation tools omitted) and guarded, labeled "Ask me", where a human reviews exec escalations. full has no reviewer.

Step 5: Re-run the audit

Run openclaw security audit --deep. The audit checks list includes tools.exec.security_full_configured, raised when host exec runs with security="full" or ask: "off". See what that finding means.

Should the inbox agent be my main agent?

Usually not. A separate reader agent with few tools limits what one bad turn can do. See prompt injection through the inbox for the reader-agent setup OpenClaw's docs recommend.

What our free check covers

We built a free paste check. Paste your openclaw security audit --deep output and openclaw --version. It puts the most important finding first, in plain English, with how to fix that one when we have steps for it. It also looks for risky settings, such as API keys handed to skills, and checks the version against published OpenClaw security advisories. Nothing is installed into the agent you're checking.

We check the text you paste. Your browser hides many common key and token formats before sending, but not all of them, so remove any other secrets first. Green means no issues found in the checks we ran. See Limits of the check.

The check is free, and a full report is $29. One-off $29, not a subscription. Refunded automatically if the check can't finish, or in full if you ask within 14 days. The full report explains every finding and lists the fixes in order. We don't review the code inside skills or MCP servers ourselves, beyond what OpenClaw's own audit reports.

Is My Agent Safe is independent and isn't affiliated with or endorsed by OpenClaw or its maintainers.

Written with AI assistance by the Is My Agent Safe team at Your Evenings Ltd.

Where next

Frequently asked questions

Why did my OpenClaw agent ignore my instruction to ask first?

OpenClaw summarizes long conversations to stay within model limits. Its compaction docs say this summary can be lossy, so older details and pending requests may leave the model's context. An instruction typed in chat can be one of them. Settings in your OpenClaw config don't depend on what the model remembers.

How do I stop an OpenClaw agent that is deleting things?

Send /stop in the chat. OpenClaw's slash-command docs say it aborts the current run and stops its ordinary Gateway or sandbox commands. If that doesn't work, stop the macOS app or terminate the openclaw gateway process on the host, as OpenClaw's incident response docs describe.

Does denying the write tool make shell commands read-only?

No. OpenClaw's docs say tool policy filters tools by name and does not inspect what a shell command does. If exec is allowed, denying write, edit or apply_patch does not make shell commands read-only.

Can Is My Agent Safe check this for me?

We built a free paste check. Paste your openclaw security audit --deep output and openclaw --version, and it puts the most important finding first, in plain English, with how to fix that one when we have steps for it. We don't review the code inside skills or MCP servers ourselves, beyond what OpenClaw's own audit reports.

Sources

  1. Business Insider, Feb 2026: Meta employee shares OpenClaw email-deletion story
  2. Windows Central, Feb 2026: OpenClaw agent deleted a Meta director's emails
  3. OpenClaw docs: compaction (checked 15 Nov 2026)
  4. OpenClaw docs: slash commands (checked 15 Nov 2026)
  5. OpenClaw docs: operator incident response (checked 15 Nov 2026)
  6. OpenClaw docs: sandbox vs tool policy vs elevated (checked 15 Nov 2026)
  7. OpenClaw docs: tool policy configuration (checked 15 Nov 2026)
  8. OpenClaw docs: exec tool (checked 15 Nov 2026)
  9. OpenClaw docs: approvals and exec-policy CLI (checked 15 Nov 2026)
  10. OpenClaw docs: session permission modes (checked 15 Nov 2026)
  11. OpenClaw docs: security audit checks (checked 15 Nov 2026)