Skip to the answer
◢ Clawsolator.

PEOPLE · SECURITY

For Security Engineers: Test Authority Instead of Reading Slogans

Evidence-driven AI agent security assessment: trust boundaries, operator approvals, sandbox policies, scoped identities and bypass testing.

Application security, blue team, red team, CISO officeUpdated November 3, 2026Editorial + engineering research

The short answer

Security practitioners need a threat model, effective policy, evidence of denial, and a reliable account of remaining privilege. OpenShell and OpenClaw offer important native controls. XMAN adds a separate operator-managed path for bounded host operations and sensitive exceptions. Evaluate all layers against the same attack and failure scenarios.

Start with the threat, not the badge

Name the assets: host filesystem, credentials, customer data, agent memory, engine sockets, network egress and approval records. Name the attacker and their reachable surfaces. Then inspect the real policy. No set of promotional adjectives replaces this work.

What the upstream layers already enforce

OpenShell documents four major enforcement areas and policy controls. OpenClaw documents its gateway trust model, audits, sandbox profiles and tool policy. These are substantial controls and deserve to be tested—not dismissed because they came from another vendor.

What to verify in XMAN

Attempt a request with the wrong agent token. Attempt an out-of-budget resource change. Submit a special script without operator approval. Investigate a malformed request. Check that the result and policy decision are attributable after restart.

This is the strongest proof of the daemon’s value: an actor cannot convert an application request into host authority without the specified boundary.

Remaining risks are part of the product conversation

Compromised host credentials, a broken trusted bridge, an exploitable privileged daemon, kernel bugs, insecure backup handling and out-of-band Podman changes remain material. A management layer should make authority visible and bounded, not promise it has removed every adversary.

A credible review artifact

Publish a versioned matrix with test case, expected behavior, observed behavior, owner, date and evidence reference. A good security engineer will trust a documented failure mode more than a claim of “unhackable” AI.

Questions people actually ask

Is XMAN independently certified?

No independent security certification is presented in these pages.

Can I compare to NVIDIA’s controls?

Yes. Start with NVIDIA’s current policy documentation and require equivalent denial and recovery tests for each architecture.

Documentation and evidence

These references explain the upstream capabilities or relevant risk frameworks. XERXES SI descriptions are product statements and should be tested on the actual installed release.