OpenClaw Hardening#

Audit the local OpenClaw setup without making assumptions about the host OS.

Guiding Principles#

Before suggesting or applying any change, explain the risk in plain terms — users can only make informed decisions when they understand what they're accepting. Changes to files, permissions, users, or startup commands require explicit user confirmation, because an agent that acts without consent removes the user's ability to course-correct.

Use only local inspection. External network scans are out of scope for a local hardening audit and can create a false sense of security by checking reachability rather than configuration intent.

When config, process state, or permissions cannot be verified, report Unable to verify rather than assuming the best case. A silent false OK is worse than an honest unknown.

Remind the user to restart OpenClaw after any accepted config change, since OpenClaw reads config at startup and changes don't take effect until then.

Audit Workflow#

1. Detect the operating environment#

Identify the platform before choosing commands.

• On Linux or macOS, prefer native shell tools such as id, ps, ss, netstat, stat, and ls.
• On Windows, prefer PowerShell equivalents such as whoami, Get-Process, Get-NetTCPConnection, Get-Acl, and Select-String.

If a command is unavailable, switch to an equivalent rather than failing the whole audit.

2. Inspect configuration sources in precedence order#

Inspect the most specific local source you can verify:

1. Running process arguments, if an OpenClaw process is already running
2. Environment variables already set in the current session
3. Local config files

Check common config locations:

• ./openclaw.json
• ~/.openclaw/config.json
• %USERPROFILE%\.openclaw\config.json

Prefer the value actually in effect. If multiple sources disagree, report the highest-precedence value and note the lower-precedence values as context.

3. Audit bind address#

Determine the effective bind or host value for the gateway.

• Treat 127.0.0.1, localhost, ::1, and loopback as secure local-only bindings.
• Treat 0.0.0.0, ::, or a concrete LAN/public IP as exposed unless the user explicitly wants remote access.
• If no bind value is set, report Secure by default if you have high confidence in the current OpenClaw version's defaults, or Unable to verify version-specific default otherwise.

If the bind address is exposed, explain that any listener on a non-loopback interface may be reachable by other devices on the network. Offer to change it to a loopback value after user confirmation.

4. Audit gateway port exposure#

Determine the effective gateway port.

• Treat 18789 as the current default when no override is configured.
• Do not assume older web-app ports such as 3000, 3001, or 8080 unless the local config or running process actually uses them.

Inspect active listeners for the effective port and pair the result with the bind audit:

• Local-only listener on loopback → secure
• Listener on 0.0.0.0, ::, or a non-loopback address → exposed
• No active listener and no running process → configuration only, not runtime-verified

5. Audit authentication mode#

Check gateway.auth.mode in the effective config.

• Flag as DANGER if the field is absent or set to anything other than "token" — missing auth means any local process can connect to the gateway.

• If "token" is set, inspect the token value without printing it back to the user:

  • Flag short tokens (< 20 characters), all-lowercase dictionary words, or values that look like placeholders ( changeme, secret, token123) as WARN.
  • Recommend storing the token via a SecretRef (env or file source) rather than inline JSON, so the credential isn't embedded in the config file itself.

6. Audit execution privileges#

Check whether OpenClaw or the current shell is running with elevated privileges.

• On Linux or macOS, flag root or uid=0 as DANGER for routine use.
• On Windows, flag an elevated Administrator session as DANGER for routine use.

Installed skills inherit the agent's privileges. Recommend a normal dedicated user account for daily operation, and provide platform-specific remediation steps only after user confirmation.

7. Audit tool execution policy#

Inspect the tools section of the effective config.

• Check tools.deny — flag as WARN if dangerous tool groups ( group:automation, group:runtime, group:fs) are not restricted for the user's stated use case.
• Check tools.exec.security — flag as WARN if not set to "deny" or "ask".
• Check tools.fs.workspaceOnly — flag as WARN if false or absent for production setups where the agent should not roam the full filesystem.
• Check agents.defaults.sandbox.mode — if Docker is available on the host, flag as WARN if sandbox mode is not enabled. Skills run in a sandbox cannot escape to the host even if compromised.

Explain that tool policy is the primary blast-radius control: a skill that exfiltrates data or deletes files can only cause harm if the tool policy allows it.

8. Audit DM and channel access policy#

Inspect channel-level access settings, particularly for publicly reachable channels (WhatsApp, Telegram, Discord).

• Check dmPolicy — flag as WARN if set to "open", since any user on the platform can then send commands to the agent.
• Check requireMention for group channels — flag as WARN if false, since the agent will respond to every group message rather than only explicit @-mentions.
• Recommend "dmPolicy": "pairing" with time-limited codes (1-hour expiry) for public-facing agents.

If the gateway is local-only with no external channel configured, mark this check OK (local only).

9. Audit secret handling#

Inspect for credential hygiene issues without printing full secret values back to the user.

Check for:

• API keys or tokens stored directly in openclaw.json or other plain config files
• Secret files with overly broad read permissions
• Accidental credential exposure in local git history, if the config directory is a git repository

Platform-appropriate permission checks:

• On Linux or macOS, flag group/world-readable files such as ~/.openclaw/config.json (recommended: 600) or a directory accessible beyond the owner (recommended: 700).
• On Windows, inspect ACLs and flag secret files readable by broad principals such as Everyone or Users.

If secrets appear in tracked history or plain config, recommend rotation and migration to environment variables or a SecretRef pointing to a local secrets file.

10. Suggest the built-in audit command#

If the openclaw CLI is on PATH, tell the user that OpenClaw ships with a built-in security audit that covers 50+ risk categories — more than this skill checks manually:

text

openclaw security audit           # standard audit
openclaw security audit --deep    # extended checks including historical config
openclaw security audit --fix     # auto-remediate safe/low-risk issues
openclaw doctor --fix             # repair config schema issues

Recommend running openclaw security audit as a follow-up step after any manual hardening.

11. Produce a concise report card#

Output a short report after the audit using plain ASCII-safe formatting:

text

OpenClaw Security Report Card
-----------------------------
[OK|WARN|DANGER|UNKNOWN] Bind Address      -> [detail]
[OK|WARN|DANGER|UNKNOWN] Gateway Port      -> [detail]
[OK|WARN|DANGER|UNKNOWN] Auth Mode         -> [detail]
[OK|WARN|DANGER|UNKNOWN] Execution User    -> [detail]
[OK|WARN|DANGER|UNKNOWN] Tool Policy       -> [detail]
[OK|WARN|DANGER|UNKNOWN] DM Access Policy  -> [detail]
[OK|WARN|DANGER|UNKNOWN] Secret Hygiene    -> [detail]
Score: X/7
Next Step: [single highest-value action]

Score conservatively:

• OK = 1 point
• WARN, DANGER, or UNKNOWN = 0 points

Decision Guidance#

• Prefer WARN over DANGER when exposure depends on user intent — for example, deliberate LAN access or a development machine where sandboxing isn't needed.
• Prefer UNKNOWN over guessing when the process is not running and config is absent.
• If the user asks for fixes, apply the smallest safe change first.
• If the setup is already well-configured across all checks, say so clearly and avoid inventing extra work.

Incident Response#

If the user suspects a skill has already compromised the installation:

1. Stop the Gateway process immediately.
2. Rotate gateway.auth.token and all provider API keys (OpenAI, Anthropic, etc.).
3. Review session logs at /tmp/openclaw/openclaw-YYYY-MM-DD.log and channel transcripts.
4. Run openclaw security audit --deep to identify residual issues.
5. Inspect SOUL.md and MEMORY.md in the agent directory for unexpected modifications — ClawHavoc attacks are known to persist by poisoning these files to alter future agent behavior.