Open-source agents, in plain English
Is OpenClaw safe?
Short answer: not out of the box, but it can be run safely enough for everyday jobs if you set it up carefully. OpenClaw runs tools, code and add-on "skills" using whatever logins you hand it, so a bad web page, email or skill can try to take it over. Security teams advise running it on a spare machine with its own accounts, keeping it private and installing only skills you trust.
Updated · By BotsDoThat
What has actually gone wrong
- A one-click takeover bug. CVE-2026-25253 let a malicious web page steal the gateway's credentials and act as you. Versions up to v2026.1.24-1 were affected; it's fixed, so update (depthfirst write-up).
- Malicious skills. Kaspersky found more than 600 malicious skills from 24 accounts on the skill hub in April 2026. Skills had no security check before February 7, 2026; scanning has been added since, but attacks continue (Securelist).
- Prompt injection. Any web page, email or file the agent reads can carry hidden instructions. OpenClaw's FAQ says this matters even if you're the only person who messages the bot (OpenClaw FAQ).
6 rules for running OpenClaw safely
- Run it on a spare machine. Microsoft's security team says not to run OpenClaw on a normal work or personal computer, and to use a dedicated virtual machine or separate device instead.
- Give it its own accounts. Use accounts and keys made just for the agent, not your main email or bank logins, and rotate them.
- Keep the gateway private. OpenClaw's own baseline: gateway bound to loopback (your machine only) or behind authenticated private access, DMs in pairing or allowlist mode.
- Treat skills as code. Install only skills from authors you trust, read them first, pin versions. OpenClaw says ClawHub scans are not a complete security boundary.
- Lock down risky tools. Deny or sandbox exec, browser and cron for any agent that reads web pages or email from strangers. Use a read-only reader agent for untrusted content.
- Keep it updated, then audit. Update OpenClaw and run openclaw security audit --deep after config changes.
Sources: Microsoft Security Blog, OpenClaw security FAQ, OpenClaw security docs, sandboxing.
Who should (and shouldn't) use it
If you're comfortable with a spare computer or a small server, separate accounts and reading config files, OpenClaw is a powerful free agent. If you'd rather not manage any of that, a hosted agent app with built-in approvals is the simpler choice: see what an AI agent is and is it safe to give an AI agent your email and calendar? Also useful: keep an AI agent from deleting your files and set spend limits and usage alerts on your AI accounts.
Questions
Is OpenClaw safe to use?
It can be, but not out of the box. OpenClaw is open-source software that runs tools and code with the logins you give it. Microsoft's security team says to treat it as untrusted code execution with persistent credentials and run it only in an isolated machine with its own accounts. Keep the gateway private, install only trusted skills, and keep risky tools sandboxed.
Has OpenClaw had security problems?
Yes. CVE-2026-25253 let a malicious web page steal gateway credentials and control the agent; it affected versions up to v2026.1.24-1 and was fixed (depthfirst). Kaspersky reports finding more than 600 malicious skills from 24 accounts on the skill hub in April 2026, and says skills had no security check at all before February 7, 2026 (Securelist).
Are ClawHub skills safe?
Not automatically. OpenClaw's own FAQ says to treat third-party skills and plugins as code you are choosing to trust: scans are shown before install but are not a complete security boundary. Prefer trusted authors, pin versions and read a skill before enabling it.
Is Hermes Agent safer than OpenClaw?
They make different default choices rather than one being safe and one not. Hermes asks before dangerous commands by default and its email channel ignores unknown senders unless you change that (Hermes docs). Either one is only as safe as the access you give it. Compare them: OpenClaw vs Hermes Agent.
Should I run OpenClaw on my main computer?
Security teams say no. Microsoft recommends a dedicated virtual machine or separate device that holds no sensitive data, treated as disposable, with a plan to rebuild it.
Next step
Choosing between the two big open-source agents? Read OpenClaw vs Hermes Agent, or compare others in OpenClaw alternatives for beginners. Worried about the bill? Switch OpenClaw or Hermes to a cheaper model.
Sources
- Microsoft Security Blog (Feb 19, 2026): Running OpenClaw safely
- Kaspersky Securelist: Security risks for OpenClaw users
- depthfirst: 1-click RCE in OpenClaw (CVE-2026-25253)
- OpenClaw FAQ: Security and access control
- OpenClaw docs: Security
- OpenClaw docs: Prompt injection
- OpenClaw docs: Sandboxing
- Hermes docs: Security