DEV Community

Dar Fazulyanov
Dar Fazulyanov

Posted on • Originally published at clawmoat.com

Microsoft Says Don't Run OpenClaw on Your Workstation. Here's How to Do It Safely.

On February 19, Microsoft's security team published a blog post that should make every OpenClaw user pause: "Running OpenClaw safely: identity, isolation, and runtime risk."

Their recommendation? Don't run it on your workstation at all.

"OpenClaw should be treated as untrusted code execution with persistent credentials. It is not appropriate to run on a standard personal or enterprise workstation."
— Microsoft Security Blog, February 19, 2026

They're not wrong. But their solution — spinning up dedicated VMs for every agent — isn't practical for most teams. We built a better answer.

What Microsoft Found

Microsoft identified three risks that materialize "quickly" in unguarded OpenClaw deployments:

  1. Credential exposure. Your agent can read SSH keys, AWS tokens, browser cookies, and API secrets — and exfiltrate them through a single curl command.
  2. Memory poisoning. An attacker can modify your agent's persistent state, causing it to follow malicious instructions across sessions.
  3. Host compromise. The agent can be induced to download and execute malicious code.

The Numbers Are Stark

  • 135K+ exposed OpenClaw instances (SecurityScorecard)
  • 341+ malicious skills found on ClawHub (Snyk — 13.4% critical)
  • CVE-2026-25253 scored CVSS 8.8
  • Runlayer's security team compromised an agent in 40 messages, one hour

Microsoft's Recommendation: Isolate Everything

Microsoft recommends:

  • Dedicated VM or separate physical system
  • Dedicated, non-privileged credentials
  • Access only to non-sensitive data
  • Continuous monitoring and a rebuild plan

This is sound advice. It's also wildly impractical for most people running OpenClaw.

Most users are developers on their laptops. Small businesses on a Mac Mini. Solopreneurs with one machine. They're not spinning up VMs.

The Gap Nobody's Filling

Here's what other tools cover:

Tool What It Protects Host-Level?
LlamaFirewall (Meta) Model from prompt injection
NeMo Guardrails (NVIDIA) Conversation boundaries
Runlayer Enterprise governance (SaaS) Partial
KiloClaw (Kilo.ai) Managed hosting ❌ (cloud only)
ClawMoat Your file system, credentials, shell

None of them put a security layer between the agent and your SSH keys.

How ClawMoat Fills the Gap

ClawMoat is the only open-source tool designed for host-level agent security:

Your Agent → ClawMoat (validate every action) → Your Machine (restricted)
Enter fullscreen mode Exit fullscreen mode

Four Permission Tiers

  • Observer — read-only
  • Worker — safe commands (git, npm). No destructive ops.
  • Standard — most operations. Forbidden zones enforced.
  • Full — unrestricted. Forbidden zones still active.

Forbidden Zones (Even at Full Tier)

~/.ssh/          # SSH keys
~/.aws/          # AWS credentials
~/.gnupg/        # GPG keys
~/.kube/         # Kubernetes configs
~/Library/Cookies/  # Browser sessions
# 20+ patterns total
Enter fullscreen mode Exit fullscreen mode

Built-In Monitoring

  • Full audit trail of every file access and shell command
  • Credential file monitoring
  • Network egress logging with domain blocklists
  • Real-time alerts (webhook, Slack, email)
  • Skill integrity checking

One Command

npm install -g clawmoat
Enter fullscreen mode Exit fullscreen mode

Zero dependencies. Sub-millisecond validation. MIT licensed.

Microsoft's Advice → ClawMoat Equivalent

Microsoft Says ClawMoat Does
Dedicated VM Permission tiers + forbidden zones (no VM needed)
Non-privileged credentials Worker tier blocks credential access
Non-sensitive data only 20+ forbidden zone patterns
Continuous monitoring Full audit trail + real-time alerts
Rebuild plan Incident forensics from logs in 30 seconds

The Ecosystem Is Growing Fast

In the past week:

  • Runlayer launched "OpenClaw for Enterprise" (VentureBeat)
  • Crittora announced crypto policy enforcement (Yahoo Finance)
  • KiloClaw launched managed hosting on Fly.io (VentureBeat)
  • Forbes called the OpenAI acquisition a win for small business
  • OpenClaw hit 161K+ GitHub stars

The market has spoken: agents are here to stay. The question is how to use them safely.

Get Started

npm install -g clawmoat
Enter fullscreen mode Exit fullscreen mode

ClawMoat is open source (MIT). 142 tests. Zero dependencies. They protect the model — we protect your machine.

Top comments (0)