OpenClaw Browser Control Risks: How to Limit What the Agent's Browser Can Reach
Published 2026-11-02 · Updated 2026-11-01 · By The Is My Agent Safe team
Short answer
OpenClaw's docs say enabling browser control gives the model a real browser, and if that profile is already logged in, the model can access those accounts and data. Keep the agent on the isolated openclaw profile, sign it into only what it needs, set browser.evaluateEnabled to false if you don't need page JavaScript, leave private-network blocking on, keep allowHostControl off for sandboxed agents, and keep browser control off the LAN and the public internet. Then run openclaw security audit --deep.
- Profiles: openclaw is the isolated default; user and chrome drive your real signed-in Chrome.
- Page JavaScript: browser.evaluateEnabled defaults to true; set it to false if you don't need it.
- Private network: blocked by default unless you set dangerouslyAllowPrivateNetwork to true.
- Sandboxed agents: agents.defaults.sandbox.browser.allowHostControl defaults to false; keep it that way.
- Remote: set gateway.nodes.browser.mode to off when you don't need browser proxy routing, and keep hosts tailnet-only.
Run the free check: after the changes below, run openclaw security audit --deep and paste the output into our free check (Limits of the check).
Who this is for
Anyone who has turned on, or is about to turn on, OpenClaw's browser tool. That includes small businesses and the agencies that set OpenClaw up for them.
The risk in one line
OpenClaw's browser control risks page says:
The same page says that for remote gateways, "browser control" is equivalent to "operator access" to whatever that profile can reach. A web page the agent reads can carry instructions too, so what the browser can reach is what a prompt injection could reach.
Step 1: Decide whether the agent needs a browser
The browser setup docs say the default browser tool is a bundled plugin, and that it needs both plugins.entries.browser.enabled and browser.enabled=true. The configuration docs give browser.enabled a default of true.
If an agent only needs to search and read pages, the setup docs note that tools.profile: "coding" includes web_search and web_fetch, but not the full browser tool. If no agent needs the browser, set:
{
browser: {
enabled: false,
},
}
Step 2: Keep the agent on its own profile
The browser profiles docs list three built-in profiles:
| Profile | What it drives |
|---|---|
openclaw | "managed, isolated browser (no extension required)" |
user | your "real signed-in Chrome session" through Chrome DevTools MCP; Chrome shows an "Allow remote debugging?" prompt the first time, "so someone must be at the computer" |
chrome | your "real signed-in Chrome session" through the OpenClaw browser extension; it "works from a phone with nobody at the desk" |
The configuration docs say the default profile is openclaw, and defaultProfile: "user" opts into the signed-in user browser. To pin the default, set browser.defaultProfile: "openclaw".
The existing-session docs say attaching to your signed-in Chrome "is higher-risk than the isolated openclaw profile because it can act inside your signed-in browser session." The browser control risks page adds that existing-session mode "can act as you in whatever that host Chrome profile can reach."
If you do pair the Chrome extension, the same page says its access mode is stored in Chrome, not in Gateway config. "All tabs" exposes every eligible ordinary tab in that Chrome profile except session-paused tabs, and "Selected tabs" uses the OpenClaw tab group as its access list.
Step 3: Treat the agent's profile as sensitive
The browser control API docs say: "The openclaw browser profile may contain logged-in sessions. Treat it as sensitive." The browser control risks page says to:
- avoid your personal daily-driver profile;
- "Disable browser sync/password managers in the agent profile if possible";
- "Treat browser downloads as untrusted input; prefer an isolated downloads directory."
When a site needs a login, the browser login docs recommend signing in manually in the host browser's openclaw profile. Sign it into only the accounts the agent needs, and give it its own accounts where you can. See OpenClaw VPS security for keeping the agent's accounts separate.
Step 4: Turn off page JavaScript if you don't need it
The browser control API docs say browser act kind=evaluate, openclaw browser evaluate and wait --fn "execute arbitrary JavaScript in the page context. Prompt injection can steer this. Disable it with browser.evaluateEnabled=false if you do not need it." The configuration docs give its default as true.
{
browser: {
evaluateEnabled: false,
},
}
Step 5: Keep private-network blocking on
The browser control risks page says private and internal destinations stay blocked by default: browser.ssrfPolicy.dangerouslyAllowPrivateNetwork is unset, and setting it to true allows them. Leave it unset, or don't set it to true.
The configuration docs say setting it to false explicitly turns on stricter checks: "http(s) navigation requires an IP address in the URL unless the hostname is listed in allowedHostnames." browser.ssrfPolicy.blockedHostnames denies exact hosts and wildcard subdomains before any allow rule. The docs note that *.example.com does not block example.com itself, so add both.
One default to know about: the browser security docs say that when browser.ssrfPolicy is absent, agents with unrestricted execution on the Gateway host can open localhost, 127.0.0.1 and ::1 on any port in OpenClaw's own managed Chrome, including unrelated services bound to loopback. Any explicit browser.ssrfPolicy, including {}, turns that exception off.
This policy has limits. The browser control risks page says page routing "remains request-level interception, not a network firewall", and that "complete prevention requires owner-side egress isolation or a policy-enforcing proxy."
Step 6: Keep sandboxed agents off the host browser
The browser control risks page says: "Keep host browser control disabled for sandboxed agents unless you trust them." The Docker backend docs say agents.defaults.sandbox.browser.allowHostControl (default false) lets sandboxed sessions target the host browser explicitly. Leave it at false.
Step 7: Lock down remote browser access
From the browser control risks page:
- "Keep Gateway and node hosts tailnet-only; avoid exposing browser control ports to LAN or public internet."
- "Disable browser proxy routing when not needed (
gateway.nodes.browser.mode="off")." - Treat node pairing like admin access.
browser.extensionRelay.allowLegacyAuthdefaults totruefor one migration window. "Update every relay client, then set it tofalse."
The browser security docs add, for remote CDP: prefer encrypted endpoints (HTTPS or WSS) and short-lived tokens, avoid embedding long-lived tokens directly in config files, and treat remote CDP URLs and tokens as secrets. See giving AI agents API keys without exposing them for where to keep them.
Step 8: Check the result
Run openclaw security audit --deep. OpenClaw's audit checks reference includes these browser findings:
| Check ID | Severity | What it means |
|---|---|---|
browser.control_no_auth | critical | Browser control exposed without token/password auth |
browser.remote_cdp_http | warn | Remote CDP over plain HTTP lacks transport encryption |
browser.remote_cdp_private_host | warn | Remote CDP targets a private/internal host |
browser.extension_relay_legacy_auth | warn | Legacy extension relay bearer/Basic/token authentication remains enabled |
The security CLI docs also say the audit warns when small models (<=300B parameters) are used without sandboxing and with web/browser tools enabled. For how to read the rest of the output, see OpenClaw security audit explained.
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.
Setting this up for clients? See Is My Agent Safe for agencies.
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
- OpenClaw Security Audit: What Each Finding Means and What to Fix First
- OpenClaw security checklist for client deployments
- OpenClaw VPS security
Frequently asked questions
Can the OpenClaw agent use the accounts I'm logged into?
Yes, if its browser profile is logged into them. OpenClaw's browser control risks page says that if the profile already has logged-in sessions, the model can access those accounts and data, and to treat browser profiles as sensitive state.
Which OpenClaw browser profile should I use?
The isolated openclaw profile, which is the default. OpenClaw's docs say the user profile attaches to your real signed-in Chrome and the chrome profile drives it through the OpenClaw extension, and that existing-session attach is higher-risk than the isolated openclaw profile.
What does browser.evaluateEnabled do?
It controls whether the agent can run arbitrary JavaScript in the page. OpenClaw's browser control API docs say prompt injection can steer this, and to disable it with browser.evaluateEnabled=false if you don't need it. The configuration docs give its default as true.
Does OpenClaw's browser SSRF policy stop every private-network request?
No. The docs say private destinations are blocked by default, but page routing is request-level interception, not a network firewall, and complete prevention requires owner-side egress isolation or a policy-enforcing proxy.
Sources
- OpenClaw docs: browser control risks
- OpenClaw docs: browser profiles
- OpenClaw docs: multi-profile and existing-session attach
- OpenClaw docs: browser setup
- OpenClaw docs: browser configuration
- OpenClaw docs: browser control API (security and privacy)
- OpenClaw docs: browser security
- OpenClaw docs: browser login
- OpenClaw docs: Docker sandbox backend (sandboxed browser)
- OpenClaw docs: remote and hosted browsers
- OpenClaw docs: security audit checks
- OpenClaw docs: security CLI
Get the plain-English agent security brief
Agents & AI Security, explained: a short, plain-English email when we have something useful. Free. Unsubscribe any time.
We'll only use your email to send the newsletter. The newsletter is sent by beehiiv, which records opens and link clicks. Privacy policy.