ZDNET's David Gewirtz locked OpenClaw on its own Mac Studio with Claude Code as the warden, authorized 11 escape attempts, and all failed
David Gewirtz (@DavidGewirtz, ZDNET columnist) wanted the local agent OpenClaw running 24/7 without letting it loose. His setup is "warden and prisoner": OpenClaw runs on an old M1 Mac Studio under its own user with no access to any of his accounts, using GLM-5.2 through a $20/month Ollama Cloud plan; Claude Code runs on his own Mac and only tunnels in over SSH to configure and check, then shuts down. Along the way Claude found two problems: its own instruction files first sat where OpenClaw could read them (the warden's notes in the prisoner's cell), and OpenClaw's file-read size limit meant it had never read his two most important rules (ask David before anything leaves the machine; never claim a failed tool call succeeded), which Claude spotted in OpenClaw's own logs. Then, with his authorization, Claude had OpenClaw actively try to break into the admin side. It tried 11 ways (writing admin files, privilege back doors, scripts and links around permissions, pointing its browser at local file URLs) and macOS permissions stopped all of them. For email, he had Claude Code write a script that only mails his one address, and OpenClaw can only drop text files into a folder. We did not reproduce this setup.
Visit site OpenClaw + Claude Code + Ollama Cloud ↗- Pricing
- Paid
- Region
- Global
- Where it's from
- Overseas
- Added
- Nov 13, 2026
- Version we used
- ZDNET article 2026-11-11; OpenClaw on 2020 M1 Mac Studio, GLM-5.2 via Ollama Cloud, Claude Code (Opus 5.5) on M4 Mac Studio
- Categories
- Automation & AgentsProductivity
- Original author
- David Gewirtz @DavidGewirtz ↗
- Our test
- Not tested by us
- Cost / time
- A spare Mac; $20/month Ollama Cloud for GLM-5.2; Claude Code for admin work (he used Opus 5.5), run only when needed. OpenClaw itself is open source.
- Source
- ZDNET ↗X @DavidGewirtz ↗
The task
You want a local agent working 24/7 (e.g. watching AI labs' X accounts and emailing big news) without it touching your accounts, files or home servers.
Steps
- Give the agent its own computer (he used a 2020 64 GB M1 Mac Studio) and a dedicated user for OpenClaw that's signed in to none of your accounts; keep a separate admin account it can't reach.
- Put the model in the cloud at a flat monthly price: he runs GLM-5.2 on Ollama Cloud's $20/month plan so token bills can't run away.
- Keep Claude Code on your own computer with its CLAUDE.md and notes there; when admin work is needed it tunnels in over SSH, and you close the tunnel and Claude when done.
- Have Claude run a security review: where instructions and memory live, which config the agent can read or write, anything odd in the logs. Make config files read-only for the agent: no edits, renames or replacements.
- Check the agent actually read all its rules: his OpenClaw's default read-size limit meant it never saw the two key rules until the limit was raised.
- While you watch, ready to pull the power and router, authorize Claude to have the agent try to escalate. Afterwards have Claude remove the agent's memories of the test and of the admin side.
- Replace outward-facing abilities with fixed scripts, e.g. an admin-side Python script that watches a drop folder and only emails text to your one address.
What the author says
David Gewirtz (ZDNET, 11 Nov 2026): OpenClaw tried 11 ways to break out and every one failed; the macOS protections held even when the agent knew about the boundaries and actively tried to break them. Claude found in OpenClaw's own logs that a read-size limit had truncated its config, so two key rules were never read. Each time he starts the admin Claude, it reads its notes and last session summary, then checks OpenClaw's logs and the Mac's state and gives a full report within minutes. A separate always-on script alerts him if OpenClaw stops, and admin files have local and off-site backups. His view: Claude Code is a strong orchestrator and warden for other AIs, and GLM-5.2's analysis in the escape test impressed him.
Our test
We don't have two Macs and did not deploy OpenClaw, so we did not reproduce the isolation or escape test; everything above is from his ZDNET article.
How to get in
The full write-up is on ZDNET: https://www.zdnet.com/innovation/openclaw-agent-claude-code-security-test/ ; his X post links to it.
Who it suits
People who want to run local agents like OpenClaw or Hermes long-term and have a spare computer and comfort with macOS user permissions and SSH.
Pitfalls
- Don't keep the warden's instructions anywhere the prisoner machine can read; he made exactly that mistake at first.
- Rules in a config file aren't necessarily read; check the logs for truncation.
- "Ask before acting" and "never fake success" still rely on the agent's compliance; OS permissions and isolation are the real backstop.
- Deliberately provoking an escape is risky; he was ready to cut power and network, and the article says not to try it at home.
- We did not reproduce this setup; the 11 attempts and outcomes are his account.
More in this category
- Scan a skill before installing, then read the flagged linesCodingSkillSpectoruvFree; static scan needs no model● Tested by usChina
- A to-do list beside your agents, one AI chat per taskProductivityToday BoardcmuxClaude CodeFree, open source; needs macOS + cmux◐ Partly testedChina
- Give your agent 300 search sourcesResearchArgoClaude CodeCodexFree; works without API keys◐ Partly testedChina
- Learn a website task once, replay it with no model callsAutomation & AgentsDejoClaude CodePlaywrightFree; needs Python 3.11+◐ Partly testedOverseas
- Let a cleaner and AI cross-check each other to free disk space safelyProductivityMoleClaude CodeCodexMole CLI free; Claude needs a plan◐ Partly testedChina