OpenClaw's current trust model
The official project defines the trusted operator and in-process plugin boundaries, alongside native execution approval and sandbox controls.
Security is not a tagline. It is the mapping between a request, an identity, permitted resources, reviewed changes and the evidence that verifies the result.
XMAN 0.4.76a1 is an alpha manager, not a replacement for kernel, container, Gateway, or cloud security updates. The source documents a separate operator path; individual deployments still require acceptance tests.
The controls below are separated by upstream responsibility, XMAN mechanism and remaining qualification. They are not claimed to have eliminated every possible exploit.
| Risk or boundary | Underlying failure | Clawsolator / XMAN approach | Evidence status |
|---|---|---|---|
| 01. Trusted operator per gateway | OpenClaw states the single trusted-operator scope explicitly. | Select the appropriate gateway and data boundary; do not present a single gateway as secure multi-tenancy. | Architecture decision; requires deployment verification. |
| 02. Gateway authentication and device scopes | Native gateway authenticates and scopes requests; CVEs show why patches matter. | Prefer authenticated access through a controlled boundary; XMAN cannot override native auth defects. | Upstream patch + configuration. |
| 03. In-process plugin privileges | OpenClaw notes plugins run as trusted Gateway code. | Reduce installed plugins; reviewed skill import and host-side XMAN service for narrowly brokered requests. | Layered mitigation; not plugin sandboxing. |
| 04. Host execution and shell access | Native tooling offers exec policies and optional sandboxing. | No raw Podman socket to the agent; XMAN requests are typed, bounded and reviewed. | Documented in XMAN 0.4.76a1. |
| 05. Sandbox opt-in | OpenClaw documents that sandboxing is opt-in. | Rootless Podman isolation and per-session containers when configured. | Runtime acceptance test required. |
| 06. Privileged scripts | Native tools support approvals; older versions had bypass CVEs. | Exact immutable scripts, reviewed in a restricted viewer, executed by an operator; no automatic replay after crash. | XMAN source-documented; host test required. |
| 07. Resource limits | One noisy agent can exhaust memory/CPU without appropriate limits. | Per-container CPU/RAM ceilings, aggregate envelope, reserve and Podman inspect readback. | Documented; actual kernel/cgroup enforcement needs qualification. |
| 08. Sensitive directories | A sandbox with broad mounts still reads broad files. | Least-privilege mounts and private per-agent data; separate protected domains. | Configuration-specific. |
| 09. Network egress | Any reachable destination may become an exfiltration path. | Explicit network/mount operator handoffs and reviewed policies; no claim of universal deny-by-default egress today. | Additional verification; NemoClaw already offers strong egress controls. |
| 10. Prompt injection | Untrusted content can direct models toward unintended actions. | Read-only research roles, tool restrictions, human review, approvals outside the model. | Layered mitigation; not a prompt-injection cure. |
| 11. Credential handling | Tokens in prompts/logs/files can leak with tool execution. | Private operator storage, bounded logging, no raw secrets in demos. | XMAN handles its own secrets; broader app-level credential review needed. |
| 12. Durable decisions and crashes | An approved task must not vanish or repeat after restart. | SQLite ledger, receipts, expiry, idempotency and crash-uncertainty investigation before script replay. | XMAN 0.4.76a1 documented. |
| 13. Agent-to-agent coordination | Cross-agent access risks are driven by shared trust/paths. | Explicit agent roles and bounded master help inbox; no permission grant via chat alone. | XMAN source-documented. |
| 14. OS-level kernel vulnerabilities | Rootless containers do not automatically defend against all kernel/host vulnerabilities. | Keep OS, OpenClaw, Podman and dependencies patched; verify actual runtime boundaries. | Upstream security and operator responsibility. |
| 15. Action and evidence visibility | An appealing UI can conceal what actually executed. | Show before/after, operator action, Podman readback and limits; disclose simulated examples. | Design principle; not all deployments qualified. |
The official project defines the trusted operator and in-process plugin boundaries, alongside native execution approval and sandbox controls.
Read the packaged 0.4.76a1 source, README and tests for exact-script review, signed-off operator actions, ledger continuity, per-container resource budgets and qualification limits.