OpenClaw Security: Securing a Self-Hosted Fleet of AI Agents

- Share:





3366 Members
TL;DR: OpenClaw is an open-source personal agent that runs on your own machine and can browse, fill forms, read and write files, run shell commands, and message people for you. Its security page says it assumes "one trusted operator running multiple agents per gateway," and it treats prompt injection on its own, and plugins the operator installed, as outside its vulnerability scope. That is a fair model, and it puts the boundary between your agents in your config. Add policy outside the agents and their skills: an identity per agent, deny by default, a check on every call, a human at risky steps, and one decision log. Software Factory Security applies those controls to fleets of agents.
OpenClaw is an open-source assistant that "organizes your inbox, sends emails, manages your calendar, checks you in for flights," all from the chat apps you already use: WhatsApp, Telegram, Discord, Slack, Signal, iMessage, and more. It runs on Mac, Windows, or Linux, keeps persistent memory, and supports community skills and plugins (it "can even write its own"). It is MIT licensed and stewarded by the OpenClaw Foundation, a US 501(c)(3).
The site is direct about reach: "Full System Access: Browses the web, fills forms, reads and writes files, runs shell commands. Full access or sandboxed, your choice."
The GitHub README explains the shape: "One Gateway runs it as a personal claw on a laptop or as a shared multiplayer claw." The Gateway is the local control plane for sessions, tools, events, and channels.
OpenClaw's docs give practical advice, and they are honest about limits.
allowFrom list, a dedicated phone number, and reviewing proactive automation jobs. It warns that linking your personal WhatsApp makes every message to you agent input.OpenClaw has also had real incidents. CVE-2026-25253 let a malicious page exfiltrate the gateway token through an unvalidated gatewayUrl parameter, enabling one-click compromise; it was fixed in 2026.1.29. Immersive Labs investigated malicious skills published to ClawHub, which is the supply chain risk any plugin ecosystem carries.
This line from the security page deserves attention: OpenClaw assumes "one trusted operator running multiple agents per gateway, not a shared multi-tenant service." Under that model, the following are not vulnerabilities by themselves:
That is a reasonable scope for an open-source project. It also tells you where responsibility lands. If an agent is talked into misusing access it legitimately has, or a skill you installed does something you did not expect, OpenClaw considers that your policy decision. A fleet on one gateway is a fleet, and the separation between its agents is whatever your configuration says.
| Fleet risk | How it shows up on one gateway | What fixes it |
|---|---|---|
| Standing access | Tools run on the host unless sandboxed | Deny by default per agent and workflow |
| Delegated identity | Downstream systems see your accounts | An identity per agent, bound to you |
| Intent drift | Proactive automations run while you sleep | Declared intent checked on each call |
| Prompt injection | Messages, web pages, and skill output steer the agent | Policy outside the agent and its skills |
| Audit | Each gateway and each vendor keeps its own records | One decision log |
A skill runs with the agent's authority, and the agent runs with yours. Config allowlists and sandboxes are useful, and you should use them, but they live next to the thing they constrain. If the agent can edit its own workspace or write a new skill, it is close to its own rules.
Software Factory Security makes the same argument for coding agents. As shared in our docs, "Guardrails in a prompt are suggestions. Policy outside the model is a guarantee." The controls are: give every agent its own identity delegated from a human, deny by default, check every call against policy outside the model, stop risky steps for a person, and keep one decision log.
For a self-hosted fleet, that means putting the decision at the systems the agents reach, not inside the gateway:
Zero standing privileges is the principle behind deny by default: no agent holds access it is not using for the current task. And when something goes wrong, the agent's account of what it did is not evidence. We explain why in you can't audit an AI agent by asking it. The full architecture is in the Software Factory Security blueprint.
allowFrom on every channel and use a dedicated number, as the setup guide recommends.Pick the agent on your gateway with host access and an inbox, and put its outbound tool calls behind a policy check and a log first. The Software Factory Security page covers the five controls.

Co-Founder / CEO at Permit.io