In short
- OpenClaw is not a chatbot with a text box. It reads your messages and files and can run commands, so anyone who steers it borrows your access.
- The Gateway listens on port 18789. On a normal host it binds to loopback. In a detected container it binds to all interfaces unless you set
gateway.bindtoloopback. - CVE-2026-25253 (CVSS 8.8) let a crafted link make OpenClaw connect to an attacker's address and send its token. It is fixed in 2026.1.29.
- Prompt injection does not need access to your bot. A web page, an email or a document the agent reads can carry instructions. OpenClaw's docs say sandboxing is opt-in.
- OpenClaw's own docs state it is not a security boundary between users who do not trust each other. One gateway, one trusted operator or team.
On this page
- What OpenClaw is, and why its security is different
- The Gateway: the door you must not leave open
- CVE-2026-25253: a link that takes the token
- Prompt injection: the content is the attacker
- Skills and plugins: other people's code, with your permissions
- Five checks to run on your OpenClaw today
- OpenClaw at work: one gateway, one trust boundary
- Where Keybrum fits, and where it does not
- Questions
- Sources
What OpenClaw is, and why its security is different
OpenClaw is an open-source personal AI agent, MIT licensed, that you run on your own computer or server. It was first published in late 2025 and was earlier known as Clawdbot and Moltbot. You talk to it through chat apps such as WhatsApp, Telegram or Slack, and it acts: it reads files, browses, sends messages and runs commands. Its repository had about 394,000 stars on GitHub in November 2026.
A chatbot that gives a wrong answer costs you a wrong answer. An agent with a shell, a browser and your logged-in accounts is a different object. Whoever controls what it does, even for one message, acts with your permissions. So OpenClaw security is less about the model and more about three boundaries: the network door, the content the agent reads, and the tools it holds.
| Boundary | The question | What goes wrong |
|---|---|---|
| The Gateway | Who can connect to the agent's control port? | A Gateway reachable from the internet, with weak or no authentication |
| Inbound content | What does the agent read that you did not write? | A page, an email or a message that carries instructions for the model |
| Tools and skills | What is the agent allowed to do, and whose code extends it? | Shell access with no sandbox, or a third-party skill that does more than it says |
| Secrets on disk | Which keys sit where the agent can read them? | API keys and tokens in the agent's workspace or its config folder |
The Gateway: the door you must not leave open
The Gateway is OpenClaw's control plane. It serves WebSocket and HTTP on one port, 18789 by default, including the Control UI. OpenClaw's documentation describes how it decides where to listen:
- On a regular host,
gateway.binddefaults toloopback: only programs on the same machine can connect. - In a detected container, when the setting is omitted it resolves to
0.0.0.0, every interface, so that port forwarding works. Whether the internet can reach it then depends on how you published the port. - Any non-loopback bind requires configured authentication. The docs say startup "fails closed without it".
Never expose the Gateway unauthenticated on 0.0.0.0.
Through 2026, several security firms published scans counting OpenClaw gateways reachable from the public internet. Their figures differ widely, so we do not quote one. The mechanism is not in dispute: a Docker command with a published port on a cloud server, and the control port of an agent that holds your accounts is on the internet. A cloud firewall that is open by default finishes the job.
CVE-2026-25253: a link that takes the token
The best-known OpenClaw vulnerability is CVE-2026-25253, published on 1 February 2026 with a CVSS score of 8.8. The record says that OpenClaw before version 2026.1.29 "obtains a gatewayUrl value from a query string and automatically makes a WebSocket connection without prompting, sending a token value".
In plain words: a user who opened a crafted link had their OpenClaw connect to an address chosen by someone else, and hand over the token that authenticates to their Gateway. The score reflects that it needs one click from the victim and no account from the attacker. It did not require the Gateway to be exposed to the internet.
Prompt injection: the content is the attacker
The second boundary is harder to close. OpenClaw's documentation is direct about it: prompt injection "does not require public DMs". Even if you are the only person who can message your agent, anything it reads can carry instructions: search results, web pages, emails, attachments, pasted logs.
The docs cite a 2026 public test of 272,000 attacks across 41 agent scenarios. Counting only cases where the agent both performed the harmful action and hid it from the user, the success rate ranged from 0.5% for the strongest model tested to 8.5% for the weakest of those listed. They add two cautions: attackers who adapt to a specific defense still succeed far more often, and smaller or older models are easier to steer. A low percentage, applied to an agent that reads hundreds of pages a week with a shell at hand, is not zero.
- Sandboxing is opt-in. The docs say so. Without it, a command the agent runs executes on the Gateway host.
- Give untrusted reading to a weaker agent. The recommended pattern is a read-only agent, with no tools, that summarises untrusted content for the main one. The docs note the summary itself stays untrusted.
- Limit the high-risk tools. Shell, browser, web fetch and web search only for the agents that need them.
- Agent-to-agent messaging is on by default. Restrict which agents may message which.
The same flaw exists in any app that puts a model behind an endpoint. If you build one, see unprotected AI endpoints.
Skills and plugins: other people's code, with your permissions
OpenClaw is extended with skills and plugins, many shared on its public registry, ClawHub. A skill is instructions and code that your agent will follow and run. Installing one is closer to installing a program than to adding a browser bookmark. OpenClaw's own threat model lists adversarial threats to ClawHub next to those against the platform.
- Read a skill's files before you install it, including any script it downloads when it runs
- Prefer skills whose source is public and whose author you can identify
- Be suspicious of a skill that asks you to paste a command into a terminal
- Keep wallets, password stores and SSH keys out of the agent's reachable filesystem
- Run agents that use third-party skills in a sandbox, with a tool policy
Five checks to run on your OpenClaw today
- 1
Run the built-in audit
openclaw security auditchecks your configuration against the project's baselines and lists findings with a severity. Start there, and fix in the order it gives. - 2
See where the Gateway listens
On the host, list the listening sockets for port 18789 with the command below.
127.0.0.1is local only.0.0.0.0or your public address means other machines can try to connect. - 3
Test from outside
From another network, such as your phone's connection, try to open your server's address on port 18789. It should not answer. OpenClaw's docs suggest a port scan of your own public address to confirm that only the ports you meant to expose are open.
- 4
Check your version
Anything before 2026.1.29 is exposed to CVE-2026-25253. Update, then rotate the Gateway token if you were on an older version for long.
- 5
Look at what the agent can read
List the files in its workspace and its config folder. Every API key there is a key a successful injection could send out. Move what the agent does not need.
# On the machine that runs OpenClaw
openclaw security audit
# Where does the Gateway listen? (Linux)
ss -ltnp | grep 18789
# Inside Docker: which host ports are published?
docker ps --format '{{.Names}} {{.Ports}}'0.0.0.0:18789-> in the Docker output means the port is published on every interface of the host.OpenClaw at work: one gateway, one trust boundary
Teams now run OpenClaw on shared servers, and this is where its design limit matters. The documentation states that OpenClaw "is not a hostile multi-tenant security boundary for mutually adversarial users sharing one agent or gateway". A gateway assumes one operator, or a team whose members trust each other. For people who should not see each other's data, the docs ask for separate gateways and credentials, ideally separate operating system users or hosts.
For a company, that turns into three rules. An employee's personal agent does not get production credentials. An agent that reads customer email does not share a gateway with one that can deploy. And an agent installed on a work laptop without anyone knowing is a case of shadow AI, to be found before it is reached from outside.
Where Keybrum fits, and where it does not
Keybrum is the automated pentest we build for web apps made with AI. It does not have an attack for the OpenClaw Gateway, and it will not tell you whether your agent's tool policy is right. Use openclaw security audit for that.
What it does cover is the app and the server next to the agent. For an owner who has proved ownership, it checks the server behind the app for admin consoles and internal ports open to the internet, and for self-hosted AI tools such as Ollama and n8n left reachable without a password. It attacks the app itself as a stranger and as two accounts. Many OpenClaw users also ship an app that their agent wrote. The case of Moltbook, a social network for AI agents whose whole database was readable with its public key, shows how that usually goes wrong.
Questions
Is OpenClaw safe to use?
It can be run safely, and its defaults on a regular install are careful: the Gateway listens locally and unknown senders must pair first. The risk comes from what you add: a published port, shell access without a sandbox, third-party skills, secrets in the workspace. Run openclaw security audit and keep the Gateway off the public internet.
What is CVE-2026-25253?
It is a vulnerability in OpenClaw before version 2026.1.29, rated 8.8 on the CVSS scale. OpenClaw took a gateway address from a link's query string and connected to it automatically, sending its token. A user who opened a crafted link could give an attacker access to their Gateway. Updating fixes it.
How do I know if my OpenClaw Gateway is exposed to the internet?
On the host, check which address port 18789 listens on: 127.0.0.1 is local only. Then try to reach the port from another network. If you run OpenClaw in Docker on a cloud server, look at the published ports, because Docker can bypass a host firewall rule.
Can OpenClaw be attacked through an email or a web page?
Yes. That is prompt injection: the agent reads content that contains instructions and follows them. OpenClaw's docs say this does not require anyone to message your bot. Strong models resist most attempts, not all, so limit the tools of any agent that reads untrusted content and use the sandbox.
Are ClawHub skills safe?
Treat a skill like any program from the internet. It runs with your agent's permissions. Read its files, prefer authors you can identify, and keep high-value secrets out of the agent's reach. OpenClaw's threat model covers attacks on ClawHub.
Can a team share one OpenClaw gateway?
Yes, if its members trust each other. OpenClaw's docs say a gateway is one trust boundary and is not designed to separate users who are adversaries. For mixed trust, run separate gateways with separate credentials.
Sources
- 1.Security (Gateway), OpenClaw Docs
- 2.Network exposure, OpenClaw Docs
- 3.Prompt injection, OpenClaw Docs
- 4.CVE-2026-25253, NIST National Vulnerability Database
- 5.openclaw/openclaw, GitHub
- 6.Hacking Moltbook: AI Social Network Reveals 1.5M API Keys, Wiz Research