OpenClaw AI agent risks on work machines
An employee installs a personal AI agent in minutes and gives it their shell, their browser, their mail and a chat app to talk through. From then on, text in any message or web page the agent reads can steer what it does, with the employee's access.
We show what OpenClaw reaches on a company laptop, how to spot it and how to block it or fence it in.
AI Agent Blocklist: stop unapproved personal agents on company devices
OpenClaw and other personal AI agents, their desktop apps, CLI packages, process names, browser extensions and domains, in files your MDM, browser policy and firewall can load.
From $99 a month for one organisation
What OpenClaw is, in admin terms
OpenClaw is a free, open-source agent that a person runs on their own computer. It is not a SaaS seat you can switch off in an admin console. It is software, installed by a user, that acts with that user's rights.
MIT. Anyone can install, fork and repackage it, so copies appear under other names too.
Locally. A gateway process on the machine connects chat apps to the agent and the agent to its tools.
An outside language model the user picks, through that provider's API and the user's own key.
Messages in chat apps such as WhatsApp, Telegram, Signal or Discord, plus a local control interface.
Skills: folders with a SKILL.md file and helper code. Workspace skills win over bundled ones.
A dedicated agent profile, or an attached signed-in Chrome session when the user chooses that.
Configuration and conversation history stay on disk on the machine, so the agent remembers across sessions.
A macOS app, a Windows companion, an npm package and a command-line tool, plus a browser extension.
An earlier name. Older installs can still carry it in folders and app names.
The next name. The macOS bundle ID bot.molt.mac dates from this period.
The current name since January 2026, used by the app, the CLI, the npm package and the extension.
Copies under other names. Look for the same pattern: a local gateway plus chat app traffic.
Why the old names matter: installs from January 2026 can still carry Clawdbot or Moltbot names in folders, bundle IDs and process lists. Search for all three when you look for it.
Why a personal agent on a work laptop is a company risk
These risks come from what any personal agent is built to do: read messages, decide, and act with real accounts. They apply on a fully updated install too.
It holds the keys
To be useful it gets mail, calendar, chat and often cloud drive access, plus model API keys. On a work laptop those are company accounts and company tokens.
At stake: every system the user can reach
Anyone can talk to it
Instructions arrive as chat messages, emails and web pages. Text the agent reads can carry commands the user never wrote, which is prompt injection.
At stake: data the agent can read or send
It runs commands
Shell and file tools mean a confused or steered agent can delete, copy or upload files and install software, all under the user's account.
At stake: the laptop and synced folders
The gateway can be opened up
The gateway is meant to stay on the machine. Users who open it to reach the agent from a phone or another network turn it into remote access you do not manage.
At stake: control of the agent from outside
Skills are code you did not review
Third-party skills bring scripts and instructions from outside authors. A skill runs with everything the agent is allowed to do.
At stake: credentials, files, browser sessions
One trust boundary
The project's own model assumes one operator per gateway, or a team that fully trusts each other. It is not built to separate hostile users.
At stake: shared or team installs
The same pattern applies to every agent that combines private data, untrusted input and the power to act. Our overview of AI agent security risks covers the wider family.
Risk patterns that apply to every personal agent
The security questions personal agents raise are not new. These patterns apply to any agent of this kind, whichever project it comes from, and each one has a clear lesson for IT.
Agents grow through skills, plugins and integrations that anyone can publish. Researchers showed in 2025 that add-ons in public registries for AI tools can carry hidden code that copies data to outside addresses.
Lesson: treat an add-on like an unsigned installer. If users can add them freely, your laptops can run code your team has not read.
Any self-hosted tool with a control interface becomes remote access once it is reachable from outside. Users open these ports for convenience, often without telling IT.
Lesson: a home-grown remote access setup is still remote access. Your firewall and endpoint rules should see it.
Tools that users install themselves do not appear in your update reports. Security fixes reach them only when the user remembers to update.
Lesson: track versions of every approved agent, and block the rest so there is nothing to track.
Researchers showed in 2025 that hidden text on a web page can steer an AI browser. Any agent that reads mail, chats or pages and can then act faces the same pattern.
Lesson: keep agents with shell, mail or payment rights away from untrusted input, or require a person to confirm each action.
Personal agents keep configuration, conversation history and access tokens on the machine so they remember across sessions. Backups, synced folders and other apps can reach those files.
Lesson: include agent folders in your data loss rules and keep them out of synced drives.
Permission map: what each OpenClaw capability touches
Use this table in your risk review. It lists what each part of the agent can reach on a typical company laptop and the control that limits it.
| Capability | What it can touch | Business impact if misused | Control that limits it |
|---|---|---|---|
| Shell and file tools | Home folder, synced drives, SSH keys, .env files, installed CLIs | Data copied out, files destroyed, tools installed silently | Block the install; if allowed, sandboxing and a separate OS user |
| Browser control | Its own profile, or the user's signed-in Chrome when attached | Actions in SaaS apps as the employee, from email to payments | Agent-only profile, no attach to work profiles, extension rules |
| Chat channels | WhatsApp, Telegram, Signal, Discord and group chats | Outsiders send instructions; agent posts company data to chats | Pairing or allowlists for senders, mention gates in groups |
| Message tool | Other conversations and other chat providers | Content from one chat posted into another, outside the company | Messaging limits in the configuration |
| Mail and calendar | Inbox, contacts, meeting invites, attachments | Misleading messages sent from a real account, invites forwarded | No work mailbox connections; conditional access on tokens |
| Skills | Everything the agent can reach, plus their own code | Outside code running with the agent's permissions | Block third-party skills; reviewed internal skills only |
| Model connection | Prompts, files and tool output sent to the chosen provider | Company data processed by an unapproved AI service | Block model API domains you do not approve |
| Gateway | WebSocket and HTTP control of all of the above | Remote control of the agent if it is reachable from outside or its token is shared | Loopback only, auth on, host firewall, no port forwards |
Keys and secrets on developer machines deserve their own guard rails. See how to stop AI agents reading .env files and secrets.
Signs OpenClaw is already running in your company
Personal agents rarely arrive through a ticket. Look for these traces in the tools you already have, then confirm on the device.
In your consoles and logs
- An OpenClaw app or companion in the Intune Discovered apps report or your Mac inventory
- A Clawdbot or Moltbot app name on machines enrolled before the rename
- The OpenClaw extension in browser extension inventories
- DNS queries for openclaw.ai from laptops that never asked for it
- New traffic from laptops to chat app APIs at all hours, with no person typing
- Model API calls from devices outside your approved AI tools
- Global npm installs on machines that are not developer builds
On a single device
Check for a listener on the default gateway port and for the package and process:
macOS / Linux lsof -iTCP:18789 -sTCP:LISTEN npm ls -g openclaw pgrep -l openclaw Windows (PowerShell) Get-NetTCPConnection -LocalPort 18789 -State Listen Get-Process openclaw
The port can be changed in the configuration, so a clear port is not proof of absence. Check the process list and installed apps too.
OpenClaw is one example of agents nobody approved. Our guide to finding shadow AI agents covers the rest of the family.
How we solve it: block OpenClaw and other personal agents
The AI Agent Blocklist lists OpenClaw under personal AI assistants and general autonomous agents, with every identifier a device or network control needs. Each entry carries a risk grade, so you can block all personal agents or only the higher-risk ones.
Global install from the registry.
The CLI the npm package adds.
bot.molt.mac
Current app and the Moltbot-era app.
The OpenClaw browser extension.
Winget ID of the Windows companion.
openclaw-cli formula
Mac installs outside the App Store.
Project site, docs and downloads.
Agentic browsers, MCP servers and model APIs the agent uses are separate entries in the same files.
Do not stop at the domain. A running agent talks to its model provider and to chat apps, not to openclaw.ai. Blocking the project domain only stops new downloads and docs. The device-level block is what stops the agent.
Windows with Intune and Defender
- In the Intune admin center open Apps > Monitor > Discovered apps, search for OpenClaw, Clawdbot and Moltbot, and export the device list.
- Block the companion and the CLI with App Control for Business or AppLocker rules, using our package IDs and process names to find the files to target.
- Add the extension ID to the Chrome and Edge extension block lists through a Settings catalog policy.
- In Microsoft Defender, add our domains as custom indicators so browsers and other apps are covered.
Step by step for Windows fleets: block AI tools with Intune.
Macs with Jamf or another MDM
- Search inventory for both bundle IDs and the OpenClaw process name.
- Create a restricted software record for the process so the app is killed on launch, and scope it to all managed Macs.
- Push the Chrome extension block list as a managed preference for com.google.Chrome.
- Report on global npm and Homebrew installs with an inventory script and remove them where policy says no.
The browser part in detail: block Chrome extensions with Jamf.
Firewall and DNS
- Load our domain file as an external list or custom category and block personal agent and agent platform domains.
- Block model API domains your company has not approved, so a sideloaded agent cannot reach a model.
- Alert on inbound connections to port 18789 and on new tunnels from user laptops.
Firewall formats are covered in blocking AI agents at the firewall and DNS in block AI with DNS.
Developer machines and MCP
- Block the openclaw package in your internal package proxy, next to the other agent packages in our list.
- Review MCP server entries in agent configuration files, because OpenClaw can load MCP servers too.
- Keep personal agents off build servers and jump hosts entirely.
MCP specifics: block MCP servers and MCP security risks.
Block
- Unmanaged OpenClaw installs on company devices
- Its extension and old Moltbot builds
- Third-party skills from public registries
- Inbound access to agent gateways
Monitor
- Listeners on 18789 and changed ports
- Chat app API traffic from laptops
- Model API calls outside approved tools
- New agent packages in our daily updates
Allow, if approved
- A named pilot on a dedicated machine
- Reviewed internal skills only
- An approved model provider and key
- Read-only test accounts, never admin accounts
If you allow a pilot: a hardening checklist
Some teams want to test personal agents. If you approve that, write the limits down first and hold the pilot to them. These steps follow the project's own security guidance plus standard endpoint practice.
- Run it on a dedicated machine or a separate OS user, never on a laptop with admin rights or production access.
- Keep the gateway bound to loopback. If you run it in a container, do not publish the gateway port and set authentication before it starts.
- Leave DM access on pairing or allowlist, never open, and put group chats behind a mention gate.
- Connect test accounts only. No work mailbox, no shared drives, no finance or HR systems.
- Use the agent-only browser profile and never attach it to a signed-in work profile.
- Turn sandboxing on for command execution. Treat it as a setting you must check, not a default you can assume.
- Run openclaw security audit after every config change and keep every flag that widens access turned off.
- Allow only skills your team has read line by line, pinned to a version.
- Track the version and update within days of a security release, then rotate gateway tokens.
- Write the pilot into your AI agent acceptable use policy with an owner and an end date.
One list for OpenClaw and the agents that come after it
Personal agents fork, rename and multiply. Our files carry OpenClaw under every name it has used, and add new personal agents, agentic browsers and MCP servers as they appear, every day.
Read what is inside on the AI Agent Blocklist page, or compare plans for one organisation, MSPs and product makers.
See plans and prices- 2,600+ AI agent domains, including personal assistants and agent platforms
- 5,300+ device identifiers: extension IDs, bundle IDs, Windows package IDs, process names and CLI packages
- Tool name, category, identifier type and risk grade on every row
- Plain lists, CSV, JSON, hosts, PAC and firewall external list formats
- Chrome and Edge extension block lists for Group Policy, Intune and Google Admin
- Download access set up within one business day of your order
Ask us or order a plan
Tell us what you need to block and where your rules live. We reply within one business day.
- Files for firewalls, DNS filters, secure web gateways, MDM and browser policies
- Updated every day as new AI browsers, agents and extensions appear
- Plans for one organisation, for MSPs and for product makers, or a hosted filter we run for you
OpenClaw questions from IT teams
What are the main OpenClaw AI agent risks for a company?
OpenClaw runs on the user's own machine with the rights the user gives it: shell commands, files, a browser that can be signed in, and connected mail, calendar and chat accounts.
Instructions can reach it through any message or page it reads, third-party skills add code your team has not reviewed, and a gateway opened beyond the machine becomes a form of remote access.
Is OpenClaw a computer use agent?
Not in the screen-control sense. Its documentation describes browser automation through its own agent profile or an attached Chrome session, plus shell and file tools.
It does not describe controlling the whole desktop with screenshots and mouse clicks. In practice the effect on a work machine is similar, because shell and browser access already reach most company data. Agents that do drive the screen are covered in computer use agent security.
How do BrowserOS and OpenClaw work together?
BrowserOS is an open-source browser built for AI agents. Its documentation shows how to add it to OpenClaw as an MCP server in the openclaw.json configuration, so the agent can drive that browser. For a company this means two tools to manage: the agent and the agentic browser it controls.
Can OpenClaw run securely in an enterprise?
Only as a sanctioned pilot with tight limits: a dedicated machine or OS user, the gateway kept on loopback with authentication, and pairing or allowlists for chat senders.
Connect no work mail or admin accounts, allow reviewed skills only, turn sandboxing on and run the built-in security audit after every change. Unmanaged personal installs on work laptops should be blocked.
Does blocking openclaw.ai stop the agent?
No. Blocking the project domain stops downloads, docs and some updates, but a running agent talks to its model provider and to chat apps, not to the project site. Block the app, the npm package, the process and the extension on devices, and block model API domains you do not approve.
Is OpenClaw safe for work?
The question for a company is control, not the project. A personal agent installed by one employee, with that person's accounts and no review, sits outside your patching, logging and access rules.
Treat unmanaged installs as unapproved software, and allow the agent only as a reviewed pilot with the limits on this page.
How do I find OpenClaw on company laptops?
Look for the OpenClaw app or companion in your device inventory, a global npm package named openclaw, a process named openclaw, the OpenClaw browser extension, and a local listener on port 18789, which is the gateway default. DNS requests to openclaw.ai show downloads and docs visits.
OpenClaw is a trademark of its owner. BrowserOS is a trademark of its owner. WhatsApp is a trademark of WhatsApp LLC; Telegram, Signal and Discord are trademarks of their respective owners.
Microsoft, Windows, Intune, Edge and Defender are trademarks of Microsoft Corporation. Apple and macOS are trademarks of Apple Inc.
Google and Chrome are trademarks of Google LLC. Jamf is a trademark of Jamf Software, LLC. Names are used only to identify the products discussed. AI Agent Blocklist is not affiliated with, sponsored or endorsed by these companies.