OpenClaw security.exposure.open_groups_with_runtime_or_fs: What It Means and How to Fix It

Published 2026-10-22 · Updated 2026-10-22 · By The Is My Agent Safe team

Short answer

security.exposure.open_groups_with_runtime_or_fs is a finding from openclaw security audit. OpenClaw's docs rate it critical/warn. Follow the fix steps below, then run openclaw security audit --deep again and check that the finding no longer appears.

This page covers one finding from OpenClaw's own security audit: what it means, how to fix it, and what to check afterward. It's based on OpenClaw's docs and source, linked at the end.

Paste your audit: run openclaw security audit --deep and paste the output into our free check for a red/amber/green result with the most important finding first (Limits of the check).

How to fix it

  1. Read the finding's detail. It lists each open policy path, then each risky agent context with its sandbox mode, command tools, file tools and fs.workspaceOnly value.
  2. This stops your agent. It stays offline until you start it again a few steps below. Stop the Gateway with openclaw gateway stop before you change anything. Make a backup with openclaw backup create, and a quick copy of your config file so you can undo the change (openclaw config file shows where it is).
  3. Close each open path in the detail. For DMs, set dmPolicy to pairing or allowlist. For groups, set groupPolicy to allowlist or disabled. With allowlist, only the senders or groups you list can trigger the agent, and an empty group allowlist blocks group messages. openclaw security audit --fix switches common open group policies to allowlist, but it doesn't change DM policies or turn off tools. Closing the open policies is the only change that clears this finding for every agent.
  4. If a channel has to stay open, follow OpenClaw's own suggested fix: set tools.profile to messaging (or deny group:runtime and group:fs), set tools.fs.workspaceOnly to true, and set agents.defaults.sandbox.mode to all. Sandboxing needs a backend such as Docker, so read OpenClaw's sandbox docs first. The audit only accepts all: non-main doesn't clear it.
  5. These settings also limit the agent when you use it yourself. In 2026.10.5 the audit checks every agent, not just the one a channel is routed to, so the finding stays while any agent keeps these tools without a full sandbox and any channel is open.
  6. Start the Gateway again and run the audit. If you think someone has already used an open chat to make the agent run commands or change files, these steps aren't enough: follow OpenClaw's full incident-response plan (linked in the sources below).
openclaw gateway stop
openclaw backup create
cp ~/.openclaw/openclaw.json ~/.openclaw/openclaw.json.bak  # use the path from openclaw config file if it's different
# close each open path from the detail, for example:
openclaw config set channels.discord.groupPolicy allowlist
# or, if the channel has to stay open:
openclaw config set tools.profile messaging
openclaw config set tools.fs.workspaceOnly true
openclaw config set agents.defaults.sandbox.mode all
openclaw config validate
openclaw gateway start
openclaw security audit --deep

After any change, run openclaw config validate. If you stopped the Gateway, start it with openclaw gateway start. Then run openclaw security audit --deep and check that the finding is gone.

OpenClaw's docs say openclaw security audit --fix doesn't fix this finding, so the steps above are manual.

What this finding means

OpenClaw's audit raises this when a channel has a DM or group policy set to open and an agent can use command tools (exec, process) or file tools (read, write, edit, apply_patch) without a full sandbox. Anyone in the open chat could use prompt injection to make the agent run commands, or read and change files, on the host. In OpenClaw 2026.10.5's source, it's critical when command tools are reachable and sandbox.mode isn't all. It's warn when only file tools are reachable, and file tools don't count if tools.fs.workspaceOnly is true. The detail lists the open policy paths, then each agent context (agents.defaults and each named agent) with its sandbox mode and tools.

checkIdRating in OpenClaw's docsWhat OpenClaw's docs say it flags
security.exposure.open_groups_with_runtime_or_fscritical/warnOpen DMs/groups can reach command/file tools without sandbox/workspace guards

What it looks like in the audit output

Made-up example, not from a real agent:

CRITICAL security.exposure.open_groups_with_runtime_or_fs
Open group/DM policy with runtime/filesystem tools exposed
Found inbound policy="open" at:
- channels.discord.groupPolicy
Risky tool exposure contexts:
- agents.defaults (sandbox=off; runtime=[exec, process]; fs=[read, write, edit, apply_patch]; fs.workspaceOnly=false)
Prompt injection in open conversations can trigger command/file actions in these contexts.

False alarms and open reports

What our free check covers

The free check at /check is paste-only. We check the text you paste: your openclaw security audit output. The check never connects to your agent or tests anything live. Nothing is installed into the agent you're checking. Green means no issues found in the checks we ran. The free check puts the most important finding first, in plain English, with how to fix that one when we have steps for it. The full report explains every finding and lists the fixes in order. The check doesn't detect malicious skills. We don't review the code inside skills or MCP servers ourselves, beyond what OpenClaw's own audit reports. The check is free, and a full report is $29. See Limits of the check.

Next steps

Sources

  1. OpenClaw docs: security audit checks (checkId catalog)
  2. OpenClaw docs: openclaw backup
  3. OpenClaw docs: openclaw config (config file)
  4. OpenClaw 2026.10.5 source: src/security/audit-extra.sync.ts (where the check is defined)
  5. OpenClaw docs: tool policy (profiles, tool groups, tools.deny)
  6. OpenClaw docs: sandbox modes, scope and backend
  7. OpenClaw docs: groups (groupPolicy)
  8. OpenClaw docs: openclaw security (what --fix changes)
  9. OpenClaw docs: Gateway service commands (start, stop)
  10. OpenClaw GitHub issue #175215: exposure checks ignore exact channel-to-agent bindings (open)
  11. OpenClaw GitHub issue #43804: doctor --deep and security audit --deep disagreed (closed)
  12. OpenClaw docs: operator incident response (full plan)