audit-openclaw-security

Run a defensive, permissioned security audit of an OpenClaw deployment and turn the results into a practical remediation plan.

This revision is tuned for OpenClaw 2026.3.8 and uses {baseDir} when referencing bundled scripts from commands.

Guardrails

1. 1. Only audit systems the user owns or has explicit permission to assess.
2. Never ask for raw secrets. Do not request gateway tokens/passwords, model API keys, session cookies, OAuth creds, or raw credential files.
3. Prefer outputs that are designed to be shareable or redacted:

- openclaw status --all - openclaw status --deep - openclaw gateway probe --json - openclaw security audit --json - openclaw security audit --deep --json

1. 4. Treat the Gateway, Control UI, browser control, paired nodes, and automation surfaces as operator-level access.
2. Default to audit-only. Before any config edits, --fix operations, firewall changes, or restarts, create a backup first and get explicit user approval.
3. When the user wants remediation, make the backup step explicit:

- openclaw backup create --verify - use --no-include-workspace if the config is invalid but you still need state + creds - use --only-config if the user only wants a minimal safety copy before edits

What “good” looks like

• - Gateway is bound to loopback unless there is a deliberate, defended reason not to.
• Strong Gateway auth is enabled.
• No accidental public exposure (LAN bind, port-forward, permissive reverse proxy, Tailscale Funnel).
• Control UI is either localhost/Serve or explicitly origin-restricted behind a trusted proxy.
• DMs require pairing or strict allowlists.
• Groups require mention gating and are not open if broad tools are enabled.
• INLINECODE10 is isolated appropriately:

- per-channel-peer for most multi-user setups - per-account-channel-peer when the same provider runs multiple accounts

• - Tooling is least privilege:

- tools.profile: "messaging" or stricter for inbox-facing agents - deny group:runtime, group:fs, group:automation on untrusted surfaces - tools.fs.workspaceOnly: true - tools.exec.security: "deny" or at least approval-gated - tools.elevated.enabled: false unless there is a narrow, intentional need

• - Plugins and skills are explicitly trusted, minimally writable, and not used as an easy persistence path.
• Secrets, transcripts, and logs have tight permissions and an intentional retention plan.

Use the bundled files progressively

Only open the extra files you need for the task:

• - references/command-cheatsheet.md — exact command ladders
• INLINECODE21 — current high-signal checkId glossary
• INLINECODE23 — secure baseline snippets
• INLINECODE24
• INLINECODE25
• INLINECODE26
• INLINECODE27
• INLINECODE28 — report structure

Step 0 — Establish context quickly

Collect just enough context to choose the audit path:

• - Where is OpenClaw running?

- macOS host / Mac mini - personal laptop - Docker host - EC2 / VPS / other cloud VM

• - Install style?

- native install - Docker / Compose - source checkout

• - Do we have local shell access?

- Mode A: chat-only / user runs commands - Mode B: agent can run shell commands directly

Mode A — Assisted self-audit (chat-only)

Ask the user to run the following on the OpenClaw host and share the outputs.

Minimum audit set

CODEBLOCK0

Helpful extras

CODEBLOCK1

Safe targeted config reads

Prefer targeted reads over a full config dump:

CODEBLOCK2

DM / group follow-up checks

If the issue is “the bot is online but DMs or groups behave strangely”, check pairing and mention gating:

CODEBLOCK3

Examples of <channel> include discord, slack, signal, telegram, whatsapp, matrix, imessage, and bluebubbles.

If the user must share the config

OpenClaw config is often JSON5-like. Redact it before sharing:

CODEBLOCK4

Host / network snapshots

macOS

CODEBLOCK5

Linux / cloud VM

CODEBLOCK6

Docker / Compose

CODEBLOCK7

Mode B — Automated local audit (shell access)

Run the bundled collector and report renderer:

CODEBLOCK8

Then review openclaw-security-report.md, refine wording where needed, and present the final report to the user.

Notes on the collector

• - It is read-only by default.
• It does not run openclaw security audit --fix.
• It collects shareable CLI diagnostics plus basic host/network context.
• It now captures current high-value signals such as:

- openclaw status --deep - openclaw gateway probe --json - openclaw channels status --probe - targeted safe config get values - backup dry-run metadata

How to interpret the audit

Use OpenClaw’s own security audit output as the primary source of truth, then translate it into a clear threat narrative.

Triage order

Prioritise in this order:

1. 1. Anything open + tools enabled

Lock down DMs/groups first, then tighten tool policy and sandboxing.

1. 2. Public network exposure

LAN bind, Funnel, missing auth, weak reverse-proxy handling.

1. 3. Browser / node / Control UI exposure

Treat these as operator access, not “just another feature”.

1. 4. Filesystem permissions

State dir, config file, auth profiles, logs, and transcript locations.

1. 5. Plugin / skill supply chain

Trust only what is intentionally installed and writable by the right user.

1. 6. Model and prompt-injection resilience

Important, but not a substitute for access control.

Findings that are easy to miss in newer OpenClaw builds

Pay extra attention to these newer or high-signal check IDs:

• - INLINECODE44
• INLINECODE45
• INLINECODE46
• INLINECODE47
• INLINECODE48
• INLINECODE49
• INLINECODE50
• INLINECODE51
• INLINECODE52
• INLINECODE53
• INLINECODE54
• INLINECODE55

Use references/openclaw-audit-checks.md and assets/openclaw_checkid_map.json to map each finding to likely config paths and remediation areas.

Core remediation patterns

1) Gateway exposure and auth

• - Prefer gateway.bind: "loopback".
• Require token or password auth for anything beyond strictly local use.
• Do not treat gateway.remote.* values as protection for local WS access; actual protection comes from gateway.auth.*.
• If the user needs a new shared secret, openclaw doctor --generate-gateway-token is the safe boring path.

2) Reverse proxies and browser-origin policy

If there is a reverse proxy in front of the Gateway:

• - configure INLINECODE62
• keep gateway.allowRealIpFallback: false unless there is a very specific need
• for non-loopback Control UI use, set INLINECODE64
• do not enable Host-header origin fallback unless the user knowingly accepts the downgrade

3) Tailscale Serve vs Funnel

• - tailscale.mode: "serve" keeps the Gateway tailnet-only.
• INLINECODE66 is public and should be treated as urgent/high risk.
• INLINECODE67 can allow tokenless Control UI/WebSocket auth via Tailscale identity headers. That assumes the gateway host itself is trusted.
• If untrusted code can run on the host, or if any reverse proxy sits in front of the gateway, disable gateway.auth.allowTailscale and require token/password or trusted-proxy auth.

4) DM and group isolation

• - Use dmPolicy: "pairing" or allowlist for inbox-facing bots.
• For shared or support-style inboxes, set session.dmScope: "per-channel-peer".
• For multi-account channel setups, prefer per-account-channel-peer.
• Avoid groupPolicy: "open" unless the tool surface is extremely limited.
• Require mentions in groups and use agents.list[].groupChat.mentionPatterns where native mentions are unreliable.

5) Tool surface reduction

Start from the conservative baseline in references/openclaw-baseline-config.md.

Good defaults for user-facing agents:

• - INLINECODE76
• deny INLINECODE77
• deny INLINECODE78
• deny INLINECODE79
• INLINECODE80
• INLINECODE81 and INLINECODE82
• INLINECODE83
• INLINECODE84

6) Node / browser / automation trust

• - Paired nodes are remote execution surfaces. Audit them like you would audit operator access.
• Browser control is not “just viewing pages”; it is effectively remote operator capability.
• INLINECODE85 / cron tools create persistence and should not be reachable from untrusted chat surfaces.

7) Secrets, logs, transcripts, and writable paths

Audit and discuss these paths carefully without asking for raw contents:

• - INLINECODE87
• INLINECODE88
• INLINECODE89
• INLINECODE90
• INLINECODE91
• pairing stores under INLINECODE92

Platform-specific guidance

Load the matching playbook when the environment is clear:

• - macOS host / Mac mini -> INLINECODE93
• personal laptop -> INLINECODE94
• Docker / Compose -> INLINECODE95
• EC2 / VPS -> INLINECODE96

Deliverable format

Use assets/report-template.md or the rendered report from {baseDir}/scripts/render_report.py.

The final deliverable should include:

• - executive summary
• environment overview
• findings table with redacted evidence
• sequenced remediation plan
• verification commands
• residual risk / operational practices

Troubleshooting notes

“openclaw: command not found”

• - Confirm the CLI is installed and on PATH.
• On Windows, prefer WSL2 for shell-driven audit flows.
• Re-run the official install / update path, then retry openclaw --version.

“Gateway won’t start — configuration invalid”

OpenClaw now fails closed on invalid config keys, invalid values, or invalid types. That is intentional and security-relevant.

Use:

CODEBLOCK9

Even when the config is invalid, diagnostic commands such as openclaw status, openclaw gateway status, openclaw gateway probe, and openclaw health are still useful.

“Runtime: running” but “RPC probe: failed”

Trust the probe details, not just the supervisor status:

• - INLINECODE105
• INLINECODE106
• INLINECODE107

This often means service/config drift, auth mismatch, or a listener that is not actually reachable by the CLI.

“Bot is online but DMs fail”

Check:

CODEBLOCK10

Common root causes:

• - pending pairing approval
• INLINECODE108 too strict for the expected sender
• provider-side permission or token drift

“Groups are silent”

Check:

• - INLINECODE109
• INLINECODE110
• INLINECODE111
• audit findings about open groups combined with runtime/fs/elevated tools

Trigger tests (skill author sanity check)

Should trigger:

• - “Can you audit my OpenClaw setup for security?”
• “My OpenClaw gateway is exposed through Tailscale Serve — is that okay?”
• “Interpret my openclaw security audit --deep --json findings.”
• “I’m running OpenClaw in Docker on a VPS; help me harden it.”
• “Why is my OpenClaw Control UI complaining about origins and trusted proxies?”
• “My bot is online but DMs don’t reply; can you audit pairing and access policy?”

Should not trigger:

• - generic macOS hardening unrelated to OpenClaw
• generic Docker security unrelated to OpenClaw
• general AWS or VPS hardening unrelated to OpenClaw
• unrelated software audits