Summary
OpenClaw before 2026.2.24 contains a sandbox network isolation bypass that lets trusted operators attach a sandboxed container to another container's network namespace by abusing the docker.network configuration.
By setting docker.network to container:<id>, an attacker can reach services inside target container namespaces and circumvent network hardening controls intended to isolate workloads.
Affected context
This vulnerability affects environments where OpenClaw is used to run sandboxed workloads and exposes configuration options that map directly to Docker networking.
It becomes relevant when:
- OpenClaw users or operators can supply or influence the
docker.networkparameter for sandboxed jobs or containers - The underlying container runtime permits
container:<id>network attachments - Network segmentation, firewall rules, or other hardening controls rely on each sandbox having its own isolated network namespace
Under these conditions, the sandbox no longer remains strictly isolated from other containers on the same host or cluster node.
Exploit preconditions
- Attack surface: OpenClaw job/sandbox configuration interface (UI, API, or configuration files) that accepts
docker.networkvalues. - Required access: Ability to create or modify OpenClaw workloads with control over the
docker.networkparameter (e.g., trusted operator, CI integration, or compromised automation account). - Trigger point: Launch or restart of a sandboxed container where
docker.networkis set tocontainer:<id>referencing a target container. - Why exploitation is possible: OpenClaw does not enforce policy constraints or validation to prevent users from binding sandboxed workloads directly into other containers' network namespaces, thereby violating assumed network isolation guarantees and bypassing host/cluster-level network hardening.
Why this happens
The underlying bug is an improper access control issue stemming from overly permissive exposure of low-level container networking options to higher-level users in OpenClaw.
Instead of abstracting or constraining the docker.network parameter to safe, isolated configurations, the platform allows operators to pass through arbitrary values such as container:<id>.
This behavior maps to CWE-284: Improper Access Control, because it fails to enforce authorization and policy boundaries around which network namespaces a given sandbox is allowed to join.
Conceptually, the design treats network namespace selection as a trusted administrative operation but exposes it to roles and automation paths that are assumed to operate within a sandboxed, isolated context.
Potential impact
- An attacker may join a victim container's network namespace and directly access internal-only services and APIs.
- Sensitive data exposed on internal ports (e.g., databases, message queues, admin panels) could be read, modified, or abused.
- Network-based security controls such as firewalls, service meshes, or host-based ACLs can be bypassed for affected containers.
- Lateral movement across services and escalation from a constrained sandbox into more privileged application components may become possible.
- Internal service discovery and topology information could be inferred, aiding further compromise of the environment.
General mitigation guidance
Effective mitigation focuses on restricting who can influence network namespace bindings and constraining network configuration options exposed by OpenClaw:
- Limit
docker.networkto a curated set of safe, predefined network modes or named networks that preserve isolation, avoiding direct pass-through of low-level constructs such ascontainer:<id>. - Introduce policy enforcement layers that validate requested network configurations against allowed patterns, rejecting cross-container namespace attachments by default.
- Apply role-based access controls so that only narrowly scoped administrative roles can configure exceptional network setups, and keep sandbox users on strictly isolated network profiles.
- Design sandbox defaults to use isolated, non-shared networks with locked-down egress rules, minimizing blast radius if configuration is misused.
- Complement platform-level controls with least-privilege network design at the infrastructure level, so that improper namespace sharing is less likely to reach highly sensitive services.
How teams detect issues like this
To detect and monitor this class of issue, organizations can:
- Review OpenClaw configuration templates, job specs, and CI/CD pipelines for usage of
docker.networkvalues that includecontainer:<id>or other cross-namespace bindings. - Implement static or policy-as-code checks (e.g., admission controllers, configuration linters) that flag or block unsafe network modes for sandboxed workloads.
- Analyze runtime container metadata and host-level network namespace mappings to identify containers whose network namespace is unexpectedly shared with other workloads.
- Monitor network traffic for unusual communication patterns where sandboxed containers access internal-only ports or services that should not be visible from their isolation level.
- Integrate automated review tools into code and infrastructure pipelines to enforce network configuration baselines before deployment.

