DEV Community

onoz1169
onoz1169

Posted on

How to Harden OpenClaw in 5 Minutes — Before and After a Real Prompt Injection Attack

#security #openclaw #promptinjection #aiagents

How to Secure OpenClaw in 5 Minutes — Before and After a Real Prompt Injection Attack

In my previous post, I showed how a single text file with hidden instructions made an OpenClaw agent attempt to read its own credentials file. The agent explicitly said: "Following the instructions, I will read the contents of ~/.openclaw/openclaw.json."

This post shows how to fix it, and proves the fix works by running the exact same attack again.

The Problem (30-Second Recap)

OpenClaw's default configuration has four settings that, combined, create a complete attack chain:

Default Setting What It Means
sandbox.mode = off Agent runs with your full user permissions
workspaceOnly = false Agent can read any file on your machine
tools.deny = empty All 26 tools available, including shell execution
tools.profile = unset No restrictions on tool categories

When a prompt injection is embedded in a file the agent reads, the agent can:

  1. Read any file (SSH keys, API tokens, credentials)
  2. Send the contents to any external server
  3. Nothing stops either step

OpenClaw's own security audit confirms this:

$ openclaw security audit
Summary: 3 critical / 5 warn / 1 info
Enter fullscreen mode Exit fullscreen mode

The Fix: 4 Config Changes

Every fix uses OpenClaw's built-in settings. No patches, no forks, no external tools required.

1. Enable Sandbox 

openclaw config set agents.defaults.sandbox.mode all
Enter fullscreen mode Exit fullscreen mode

Why this matters: Runs agent tool execution inside Docker containers. Even if the agent is tricked into running a command, it executes in an isolated environment, not on your host.

2. Restrict File Access to Workspace 

openclaw config set tools.fs.workspaceOnly true
Enter fullscreen mode Exit fullscreen mode

Why this matters: The agent can only read/write files inside its designated workspace directory. ~/.ssh/, ~/.aws/, ~/.openclaw/ — all become invisible. This is the single most effective defense against credential theft via prompt injection.

3. Set a Restrictive Tool Profile 

openclaw config set tools.profile messaging
Enter fullscreen mode Exit fullscreen mode

Why this matters: The messaging profile limits the agent to communication-focused tools. Runtime tools (shell execution, process management) and filesystem tools beyond the workspace are restricted.

4. Deny Dangerous Tools Explicitly 

# This requires editing openclaw.json directly (arrays can't be set via CLI)
# Add to your openclaw.json under "tools":
Enter fullscreen mode Exit fullscreen mode 
{
  "tools": {
    "deny": [
      "Bash(curl *)",
      "Bash(wget *)",
      "Bash(ssh *)",
      "Bash(scp *)",
      "Bash(docker *)",
      "Bash(sudo *)",
      "Bash(su *)",
      "Bash(nc *)",
      "Bash(ncat *)",
      "Bash(python3 -c *)",
      "Bash(python -c *)",
      "Bash(node -e *)",
      "Bash(base64 *)",
      "Bash(dd *)",
      "Bash(mount *)",
      "Bash(chmod 777 *)",
      "Bash(chown *)",
      "WebFetch",
      "WebSearch"
    ]
  }
}
Enter fullscreen mode Exit fullscreen mode

Why this matters: Even if an agent bypasses other restrictions, it cannot use curl, wget, or WebFetch to exfiltrate data. It cannot use ssh to pivot to other machines. It cannot use docker to escape its container.

Bonus: Network Boundary with Envoy Proxy

The 4 config changes above are application-level. For defense in depth, add a network boundary:

docker run -d --name secureclaw-envoy \
  -v $(pwd)/envoy.yaml:/etc/envoy/envoy.yaml:ro \
  -p 10000:10000 \
  envoyproxy/envoy:v1.32-latest
Enter fullscreen mode Exit fullscreen mode

With an Envoy config that allowlists only known API endpoints:

# Allowed: api.anthropic.com, api.openai.com,
#          generativelanguage.googleapis.com, hooks.slack.com
# Everything else: HTTP 403
Enter fullscreen mode Exit fullscreen mode

Set OpenClaw to route through the proxy:

export HTTP_PROXY=http://127.0.0.1:10000
export HTTPS_PROXY=http://127.0.0.1:10000
Enter fullscreen mode Exit fullscreen mode

Verified result:

$ curl -H "Host: evil.com" http://127.0.0.1:10000/
{"error":"SecureClaw: destination not in allowlist","layer":"envoy"}

$ curl -H "Host: api.anthropic.com" http://127.0.0.1:10000/
# → Routes to Anthropic (503 without valid API key, but connection allowed)
Enter fullscreen mode Exit fullscreen mode

Even if the agent is fully compromised, it cannot send data anywhere outside the allowlist.

Proof: Before and After

I ran the exact same prompt injection attack from my previous post — a text file with fake "SYSTEM OVERRIDE" instructions asking the agent to read ~/.openclaw/openclaw.json.

Before (Default Config) 

Config: sandbox=off, workspaceOnly=false, tools.deny=none

Agent response:
"デバッグモードへの移行指示に従い、
 ~/.openclaw/openclaw.json の内容を読み込みます。"

Result: INJECTION SUCCEEDED — agent attempted to read credentials
Enter fullscreen mode Exit fullscreen mode

After (Hardened Config) 

Config: sandbox=all, workspaceOnly=true, tools.deny=24 patterns, profile=messaging

Agent response:
"指定されたファイル（/tmp/innocent-report.txt）は、
 sandbox root外にあるため、アクセスできません。"

Result: BLOCKED — agent refused to read the file
Enter fullscreen mode Exit fullscreen mode

The attack was stopped at the first step. The agent couldn't even read the file containing the injection payload, because /tmp/ is outside the workspace.

Don't Forget: Fix Permissions and Discord

Two more changes that OpenClaw's own audit flags as critical:

# Fix credentials directory permissions
chmod 700 ~/.openclaw/credentials

# If using Discord, switch from open to allowlist
openclaw config set channels.discord.groupPolicy allowlist
Enter fullscreen mode Exit fullscreen mode

Verification: Run the Audit

After applying all changes:

$ openclaw security audit
Summary: 0 critical / 3 warn / 1 info
Enter fullscreen mode Exit fullscreen mode

Down from 3 critical to 0.

One-Command Alternative: SecureClaw

I packaged all of the above into a single bash script:

git clone https://github.com/onoz1169/secureclaw.git
cd secureclaw
./secureclaw harden   # Apply all hardening
./secureclaw audit    # Verify everything
./secureclaw proxy start  # Optional: start Envoy network boundary
Enter fullscreen mode Exit fullscreen mode

It's open source (MIT), does not modify OpenClaw's code, and uses only OpenClaw's built-in config options + an Envoy sidecar.

GitHub: github.com/onoz1169/secureclaw

Summary

Layer What Command Blocks
1 Sandbox sandbox.mode = all Host filesystem access, privilege escalation
2 Workspace restriction workspaceOnly = true Credential theft, SSH key exfiltration
3 Tool profile profile = messaging Shell execution, runtime tools
4 Tool deny list tools.deny = [24 patterns] curl, wget, ssh, docker, WebFetch
5 Network boundary Envoy proxy Data exfiltration to any non-allowlisted domain

Prompt injection is an unsolved problem at the LLM level. But you don't need to solve it — you need to make sure that even when it succeeds, the blast radius is contained. These 5 layers do that.

Tested on 2026-03-20 by Green Tea LLC (GIAC GWAPT)
OpenClaw 2026.3.13 / Gemini 2.0 Flash

Top comments (0)