Client's OpenClaw Compromised? An Incident Runbook for Agencies

Published 2026-10-20 · Updated 2026-10-20 · By The Is My Agent Safe team

Short answer

Follow OpenClaw's own incident order: contain, rotate, audit, then write it up. Stop the client's Gateway, put it back on loopback and freeze risky DMs and groups. Keep a copy of the logs and transcripts. Rotate Gateway, remote-client and provider credentials. Review what changed, re-run the deep audit, update to 2026.10.5, and send the client a dated, secret-free note.

  • Contain first: stop the Gateway, set gateway.bind to loopback (or turn off Tailscale Funnel/Serve), disable risky DMs and remove "*" allow-all entries.
  • Keep evidence before you clean up: the Gateway log, the session transcripts and a backup archive kept where only you and the client can reach it.
  • Rotate Gateway auth, remote-client secrets on every machine that calls the Gateway, and every channel and model credential.
  • Audit, update to 2026.10.5, and reopen only after the deep audit has no unresolved criticals and the access pattern is written down.

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

Agencies, freelancers and MSPs who run OpenClaw for clients and think one of those installs has been misused: unexpected messages sent, files read that shouldn't have been, a Gateway reachable from the internet, or a provider bill that jumped. If it's your own agent and your main worry is a leaked key, start with API key leaked: the first hour.

The steps below follow OpenClaw's own operator incident response page: contain, rotate, audit, collect. We've added the parts an agency needs on top: keeping evidence, checking your other clients, and telling the client. We don't offer incident response. If client data may have been taken, or the client is in a regulated business, bring in a professional incident responder as well.

Before anything happens: agree who decides

You'll move faster if each client contract or onboarding note already says:

Your contract with the client, and any data protection rules that apply to them, may also set notice duties. Check those with your own adviser. This page doesn't cover them. This guide isn't legal advice.

1. Contain

OpenClaw's incident page gives three containment steps:

  1. Stop it. Stop the macOS app if it supervises the Gateway, or terminate the openclaw gateway process.
  2. Close exposure. Set gateway.bind: "loopback", or disable Tailscale Funnel/Serve, until you understand what happened.
  3. Freeze access. Switch risky DMs and groups to dmPolicy: "disabled" or require mentions, and remove any "*" allow-all entries.

If the client's Gateway runs under a named profile, add the global --profile <name> flag to each command, as in OpenClaw's CLI reference. Don't run commands against the default profile on a shared host by mistake.

2. Keep evidence before you clean up

Rotation and fixes change state. Take copies first, so you can still answer "what happened?" later:

3. Rotate

OpenClaw's incident page says to assume compromise if secrets leaked, and lists three groups:

  1. Gateway auth: gateway.auth.token or gateway.auth.password. Rotation applies without a restart only when the effective auth mode stays the same. Set gateway.auth.mode explicitly if you use SecretRefs. Restart after an auth-mode change or a changed process environment credential such as OPENCLAW_GATEWAY_PASSWORD.
  2. Remote client secrets: gateway.remote.token or .password, on every machine that can call the Gateway. For an agency, that often includes your own laptop and any jump host.
  3. Provider and API credentials: WhatsApp creds, Slack and Discord tokens, model API keys in OpenClaw's auth stores, and encrypted secrets payload values if you use them.

Then check nothing was left behind in plain text. openclaw secrets audit --check checks config, auth stores and legacy residues for plaintext values and unresolved references. It returns exit code 1 when --check finds plaintext and exit code 2 for unresolved references (secrets CLI).

The agency part: if any credential was reused across clients, such as one model key, one Telegram bot or one SSH key, rotate it everywhere it's used. This is why OpenClaw's trust model supports one trust boundary per Gateway. Keep credentials separate for each client too. For the provider side (new keys, spend limits, bill checks), follow API key leaked: the first hour.

4. Audit what happened

From OpenClaw's incident page:

  1. Check the Gateway logs.
  2. Review the relevant transcripts.
  3. Review recent config changes that could have widened access: gateway.bind, gateway.auth, DM and group policies, tools.elevated and plugin changes.
  4. Re-run openclaw security audit --deep and confirm critical findings are resolved.

Two points from the trust model matter when you work out who was affected:

Check your other clients too. If another client shares the host, a credential, a skill or plugin, or the same setup script, re-run the deep audit there as well.

5. Update

Once the install is contained, update it. The latest stable OpenClaw release is 2026.10.5 (October 18, 2026). If the client's version is below 2026.8.1, it's missing fixes for advisories rated high. Check the version against OpenClaw's security advisories list, including any separately versioned channel plugins. Our update guide has the backup and dry-run steps.

6. Tell the client and write it up

OpenClaw's incident page lists what to collect for a report. It also makes a good structure for your client note. The examples below are made up, to show the level of detail:

IncludeMade-up example
When it happened, the host OS and OpenClaw version"Noticed 14:10 UK on Oct 20. Ubuntu VPS, 2026.9.4, now 2026.10.5."
What the attacker sent and what the agent did"A group message asked the agent to forward invoices. It sent 3 files."
Whether the Gateway was exposed beyond loopback"Gateway was bound to LAN behind a public port. Now loopback with SSH tunnel."
What you rotated, and when"Gateway token, Telegram bot token, model key: 15:00 to 15:40 UK."
What the client needs to do"Sign in to the new key dashboard and confirm the spend limit."

Attach the redacted transcript and a short log tail. Never put live secrets in the note.

If you think you've found a bug in OpenClaw itself, email the security address on OpenClaw's trust model page and don't post publicly until it's fixed. OpenClaw's trust model says prompt-injection-only chains without a policy, auth or sandbox bypass are closed as no-action. If someone simply talked the bot into it, fix the configuration on your side: who can message it, and what tools it has.

7. Reopen only after the checks

Use the review checklist in OpenClaw's exposure runbook before the client uses the agent again:

Then save a dated openclaw security audit --deep --json file as the new baseline. Our client audit guide and weekly checks cover what comes next.

What our free check can and can't do here

The free check at /check is paste-only. Paste the client's new openclaw security audit --deep output and openclaw --version, and it explains each finding in plain English, checks for risky settings such as API keys handed to skills, and checks the version against published security advisories. Green means no issues found in the checks we ran. See Limits of the check.

It can't tell you whether an install was compromised. It doesn't read logs or transcripts, doesn't detect malicious skills, and doesn't review MCP servers or skill code. The check is free, and a full report is $29.

You can buy a founding seat on /agencies. $99/month, price locked for your first 12 paid months. After that, it moves to the then-current Agency price, with at least 30 days' email notice. Cancel any time. Offer conditions.

Where next

Frequently asked questions

What is the first thing to do if a client's OpenClaw looks compromised?

Stop it. OpenClaw's operator incident response page says to stop the macOS app (if it supervises the Gateway) or terminate the openclaw gateway process, then set gateway.bind to loopback (or disable Tailscale Funnel/Serve) until you understand what happened.

Which credentials should I rotate?

OpenClaw's incident page lists three groups: Gateway auth (gateway.auth.token or gateway.auth.password), remote client secrets (gateway.remote.token or .password) on any machine that can call the Gateway, and provider and API credentials such as WhatsApp creds, Slack and Discord tokens and model API keys. Assume compromise if secrets leaked.

Where are the logs and transcripts?

Per OpenClaw's incident page, openclaw logs (or openclaw --profile <profile> logs) shows the Gateway log. The default file is /tmp/openclaw/openclaw-YYYY-MM-DD.log, and named profiles use /tmp/openclaw/openclaw-<profile>-YYYY-MM-DD.log, unless logging.file overrides it. The same page points to transcripts under ~/.openclaw/agents/<agentId>/sessions/. The multi-agent docs say current session history is stored in openclaw-agent.sqlite in each agent's directory. The two pages differ, so check your client's install and copy both if both exist.

Should I report it to OpenClaw?

Use the security contact on OpenClaw's trust model page for a suspected OpenClaw vulnerability, and don't post it publicly until it's fixed. The same page says prompt-injection-only chains without a policy, auth or sandbox bypass are closed as no-action, so a bot that was talked into doing something is usually a configuration problem to fix on your side.

Can Is My Agent Safe tell me whether a client was compromised?

No. The free check at /check reads a pasted openclaw security audit --deep output and openclaw --version. It explains the findings, checks for risky settings and checks the version against published security advisories. It doesn't read logs or transcripts, doesn't detect malicious skills, and doesn't review MCP servers or skill code.

Sources

  1. OpenClaw docs: operator incident response (contain, rotate, audit, collect)
  2. OpenClaw docs: security trust model (one trust boundary per Gateway, reporting)
  3. OpenClaw docs: Gateway exposure runbook (rollback plan)
  4. OpenClaw docs: openclaw security (CLI)
  5. OpenClaw docs: openclaw secrets (CLI)
  6. OpenClaw docs: openclaw backup (CLI)
  7. OpenClaw docs: multi-agent routing (session storage)
  8. OpenClaw release v2026.10.5 (GitHub, Oct 18, 2026)
  9. OpenClaw GitHub security advisories
  10. OpenClaw release v2026.8.1 (Aug 31, 2026)