Guide · 2026-11-09
OpenClaw Security and Governance Guide (2026)
An AI agent that reads your messages and runs tools is closer to a new employee with shell access than to a chatbot. Security teams know how to think about that, but only if the controls are named plainly.
This guide covers OpenClaw's security model as the official docs describe it, and then the governance layer a company adds on top: who owns it, who reviews it and what gets logged. We cite the OpenClaw security docs and sandboxing docs for every platform fact. Where the docs say a limit exists, we repeat it.
Start with the trust model
The security docs state that OpenClaw has one trust boundary per Gateway. The guidance assumes a single operator, or a team whose members trust each other. It is not a hostile multi-tenant boundary. For mutually adversarial users, the docs say to split Gateways, credentials, and ideally OS users or hosts.
Governance consequence: decide who shares a Gateway before anything else. A company-wide instance for one team is fine. A shared instance for customers who must not see one another is not what the docs describe.
The docs add a second point. By default, agents with message-tool access can send across conversations and channel providers. If an agent in one channel should never post in another, configure the cross-provider restrictions on the tool-permissions page.
Know the defaults
The docs list these defaults for a regular host install.
- The Gateway binds to loopback.
- Most chat channels reply to unknown direct-message senders with a pairing code and do not process the message.
- Group access is allowlisted, usually behind a mention gate.
Container images are different. The docs say they default to an exposed bind, so pair them with authentication. Some workspace channels also trust workspace membership by default, so read each channel's page.
Governance consequence: record in your runbook which of these defaults you kept and which you changed, with a reason and an approver for each change.
Control who can talk to the agent
The docs describe four direct-message access modes, which are pairing, allowlist, open and disabled. Allowlisting has two layers, and the docs treat context visibility as separate from trigger authorization. A person who can see a conversation is not automatically allowed to trigger the agent.
Practical rules.
- Use an explicit sender list. The docs recommend
channels.whatsapp.allowFromand group mention rules, and the idea carries over to other channels. - Avoid open mode on any instance that holds company credentials.
- Require a mention in group chats so the agent does not act on every message.
- Review the allowlist when people join, leave or change roles.
Limit what the agent can do
The security docs cover control-plane tools, node execution through system.run, dynamic skills and plugins, a sub-agent delegation guardrail and a read-only mode. They also describe per-agent access profiles for multi-agent setups, and they say full access with no sandbox is an explicit option, not a default.
Two commands help. openclaw policy inspects and tests the tool policy. openclaw sandbox explain shows the effective mode, workspaces, mounts and tool policy for a session or agent.
Governance consequence: keep a short table of agents, the tools each may use and the person who approved it. Revisit it every quarter.
Sandbox tool execution
According to the sandboxing docs, sandboxing is off by default. The Gateway process always runs on the host, and only tool execution moves into the sandbox when you enable it. The docs call this "not a perfect security boundary", and describe its purpose as limiting filesystem and process access when the model misbehaves.
Supported backends are Docker, which is the default local backend, Podman, SSH, OpenShell and Crabbox. The documented minimal example looks like this.
``json5 agents: { defaults: { sandbox: { mode: "non-main", scope: "session", workspaceAccess: "none" } } } ``
Three details matter for review.
- Allow and deny tool policies apply before sandbox rules, and sandboxing does not restore a tool you denied.
tools.elevatedrunsexecoutside the sandbox. Treat any agent that has it as unsandboxed for risk purposes. To hard-disableexec, use a tool policy deny.- GitHub identity inside sandboxes is opt-in per agent. The docs note that mounting the profile read-only stops edits but not credential reads, and that
openclaw security auditwarns for each opted-in agent.
Handle secrets and prompt injection
The docs have a secrets section covering what lands on disk and which files hold credentials, including file permissions, workspace .env files, logs and transcripts, and secret scanning. Read it before you give an agent a production key.
For prompt injection, the docs say untrusted content can reach the model. They cover model choice and wrapping that bounds the risk, tell operators to keep bypass flags off in production, and advise reviewing reasoning and verbose output in group chats. Flags named dangerous* or dangerously* are where the docs place their riskiest keys.
No setting makes injection impossible. Design as if it will happen. Give each agent only the credentials it needs, and make destructive actions require a human.
Do not expose the Gateway casually
The docs ask you to check the exposure runbook before binding beyond loopback. Network exposure covers bind, firewall, Gateway auth, Tailscale, reverse proxies and dangerous flags. Browser control has an SSRF policy that is strict by default. The hardened baselines keep the Gateway local, paired and tool-restricted.
Audit on a schedule
openclaw security audit checks for drift from the defaults. The docs say its checks are cataloged by checkId, with a severity and auto-fix support, and openclaw security runs the audit, reads the findings and applies supported auto-fixes.
A workable cadence.
- Run the audit before go-live and save the report.
- Run it after every config change and every upgrade.
- Run it monthly even if nothing changed.
- Triage findings by severity, assign an owner and set a due date.
- Review auto-fixes before you accept them.
The configuration docs add a useful safety net. The schema is strict, invalid edits during hot reload are skipped, and openclaw doctor shows exact issues. Combine that with your own version control of the config file.
The governance layer
Platform controls answer "what can the agent do." Governance answers "who decided that, and how would we know if it changed."
- Ownership. One named owner per Gateway and one deputy.
- Change control. Config changes go through a ticket or pull request, never a live edit.
- Access reviews. Quarterly review of allowlists, tool policies and stored credentials.
- Logging and retention. Decide what to keep from logs and transcripts, who may read them and for how long. The docs flag that transcripts hold sensitive content.
- Incident plan. Know how to stop the Gateway, revoke the model key and rotate connected credentials. Rehearse it once.
- Vendor review. If someone else hosts or operates the instance, ask where it runs and who can see conversations.
Where outside help fits
Writing policy for a regulated environment is the slow part. AY Automate offers OpenClaw and NemoClaw enterprise setup, and its page describes custom YAML sandbox policies written to fit GDPR, HIPAA or SOC 2 requirements. It lists isolated OpenShell sandboxes, deny-by-default network egress, filesystem and process controls, a log of every policy violation, and monitoring dashboards. The page does not claim a certification, and neither do we. It quotes about two weeks to a pilot.
If you want a managed instance without the platform work, AyClaw Bots says each OpenClaw instance runs in its own isolated environment, that it does not access your conversations or train on your data, and that you can export your data and leave. Read those as vendor statements and ask for the detail in writing.
NVIDIA's NemoClaw is relevant here too. It is an open source reference stack that runs supported agents, OpenClaw by default, in OpenShell sandboxes, with baseline network rules, an operator approval flow for egress and credential handling. The repository labels it alpha, with best-effort support.
A one-page review checklist
- Gateway shared only among people who trust each other.
- Bind is loopback, or exposure was reviewed against the runbook.
- DM mode is pairing or allowlist, groups are mention-gated.
- Sandbox enabled, with a named backend and a tested tool policy.
- No
tools.elevatedwithout a written reason. - No
dangerous*keys in production config. - Secrets reviewed, key scope minimal.
openclaw security auditclean or triaged.- Config in version control, backups tested.
- Owner, deputy and incident plan named.
FAQ
Is OpenClaw safe to use in a company?
It can be, if you configure it for that. The docs give you pairing, allowlists, tool policy, sandboxing and an audit command, and they state its limits plainly. They describe a single-operator trust model and say sandboxing is off by default. Safety depends on the choices you make and review.
Does OpenClaw sandbox everything by default?
No. The sandboxing docs say sandboxing is off by default, and that even when enabled, the Gateway process runs on the host while only tool execution moves into the sandbox. Enable it in agents.defaults.sandbox and verify with openclaw sandbox explain.
Can I run one OpenClaw Gateway for customers who must be kept apart?
The docs advise against it. OpenClaw is not a hostile multi-tenant boundary, so for mutually adversarial users they say to split Gateways, credentials and ideally OS users or hosts.
Is OpenClaw or NemoClaw compliant with GDPR, HIPAA or SOC 2?
We found no certification claim in the official OpenClaw or NemoClaw pages, so we make none. Compliance is a property of your whole deployment, your policies and your audits. Sandboxing and egress controls are building blocks that help you meet those requirements.
What should I audit first?
Start with who can message the agent, then which tools it may run, then whether the sandbox is on. Run openclaw security audit to catch drift from the defaults, and review the secrets guidance before you connect production credentials.