CLAWSOLATOR ALPHABeta release planned · License under considerationRead what is verified
XERXES SI / SECURITY ENGINEERING

AI Agent Security Lattice: 15 Risk Classes, Boundaries & Evidence

The security index is organized by risk, product, evidence, and enforcement. Investigate real failure modes and see where Clawsolator aims to add accountable operations.

Independent engineering commentary by XERXES SI · Updated November 1, 2026 · Claims separated from roadmap

The principle

What makes a real security boundary?

A capable agent is not a privileged operator. An instruction is not permission. A green checkmark is not an audit.

Browse by risk

Fifteen security classes. No magical promises.

These are risk categories, not fifteen proven vulnerabilities in any individual competitor. Some historical OpenClaw advisories below are already patched. Each entry separates the problem, control objective and remaining uncertainty.

RISK CLASS 01 / Prompt injection

Prompt injection

Failure mode: Untrusted text can try to redirect model behavior.

Clawsolator / XMAN approach: Constrain actual tool permissions and execution environments; do not confuse a prompt with authorization.

Verification boundary: The available evidence does not prove that all prompt injections are detected or blocked.

Evidence: OpenClaw security status & trust model ↗ · How claims are classified

RISK CLASS 02 / Skill supply chain

Skill supply chain

Failure mode: A malicious extension can request secret access or run an attacker-selected command.

Clawsolator / XMAN approach: Approve installations, restrict paths and network reach, capture the requested action and review provenance.

Verification boundary: Automated plugin malware scanning is future roadmap; upstream scanning and updates remain necessary.

Evidence: ClawHavoc campaign summary, referencing Koi Security research ↗ · How claims are classified

RISK CLASS 04 / Host execution

Host execution

Failure mode: A connected node can execute dangerous commands under its granted OS identity.

Clawsolator / XMAN approach: Keep agent work rootless and make exceptional host actions separate, explicit, bounded requests.

Verification boundary: Rootless is not magic; mounts, exposed sockets, and user-level files remain important.

Evidence: GHSA-gv46-4xfq-jv58 – historical node approval bypass (fixed) ↗ · How claims are classified

RISK CLASS 05 / Tool and session escalation

Tool and session escalation

Failure mode: Tools or inherited session routes can cross an expected permission boundary.

Clawsolator / XMAN approach: Limit tool routing, link each request to an identity, require a review for sensitive operations.

Verification boundary: Patch affected OpenClaw versions; XMAN does not patch upstream authorization bugs.

Evidence: GHSA-943q-mwmv-hhvh – historical tool escalation (fixed) ↗ · How claims are classified

RISK CLASS 06 / Filesystem exposure

Filesystem exposure

Failure mode: A broad bind mount can make private user files reachable.

Clawsolator / XMAN approach: Use read-only mounts where practical; expose only the directories a task requires.

Verification boundary: Host file ownership and actual mount policy must be tested.

Evidence: OpenClaw rootless Podman backend ↗ · How claims are classified

RISK CLASS 07 / Network egress

Network egress

Failure mode: A confined process can still transmit accessible secrets on an allowed connection.

Clawsolator / XMAN approach: Apply destination and method policies where the runtime supports them; keep credentials out of unrestricted environments.

Verification boundary: XMAN documentation does not establish full traffic inspection or automatic DLP today.

Evidence: NVIDIA OpenShell product and policy architecture ↗ · How claims are classified

RISK CLASS 08 / Secrets in prompts and logs

Secrets in prompts and logs

Failure mode: Credentials copied into context, environment, or action output can propagate.

Clawsolator / XMAN approach: Use a secrets broker and redaction; scope the agent to what it requires.

Verification boundary: Redaction alone is not a substitute for key rotation and enforcement.

Evidence: NVIDIA OpenShell product and policy architecture ↗ · How claims are classified

RISK CLASS 09 / Human approval fatigue

Human approval fatigue

Failure mode: Unclear and repeated approval requests can train operators to click through.

Clawsolator / XMAN approach: Show proposed action, risk, scope, script, reason, and durable outcome; deny by default.

Verification boundary: Confirm implementation and UI for each deployed build.

Evidence: Hermes Agent security guide ↗ · How claims are classified

RISK CLASS 10 / Recursive agents and delegation

Recursive agents and delegation

Failure mode: A delegated agent may inherit more authority than necessary.

Clawsolator / XMAN approach: Assign named agents explicit roles and budgets, limit cross-boundary delegation.

Verification boundary: Not a proof of full cross-agent isolation without configuration tests.

Evidence: OpenClaw security status & trust model ↗ · How claims are classified

RISK CLASS 11 / CPU, memory, and API spend

CPU, memory, and API spend

Failure mode: Runaway loops can consume resources or usage budgets.

Clawsolator / XMAN approach: Resource caps, scheduling, interruption, and separate provider-side spending limits.

Verification boundary: Provider cost ceilings require provider integration; not automatically guaranteed.

Evidence: OpenClaw security status & trust model ↗ · How claims are classified

RISK CLASS 12 / Approval loss and restart

Approval loss and restart

Failure mode: An operator decision can become orphaned when a service restarts.

Clawsolator / XMAN approach: Persist request state, tie decisions to durable IDs and record restart recovery.

Verification boundary: Recovery tests belong in release evidence.

Evidence: OpenClaw security status & trust model ↗ · How claims are classified

RISK CLASS 13 / Audit forgery or gaps

Audit forgery or gaps

Failure mode: Text from a model is not reliable evidence that a tool ran.

Clawsolator / XMAN approach: Use daemon-generated request IDs, action logs, status and outcomes outside model prose.

Verification boundary: Independent log-integrity proof remains a separate verification step.

Evidence: NVIDIA OpenShell product and policy architecture ↗ · How claims are classified

RISK CLASS 14 / Managed-service visibility

Managed-service visibility

Failure mode: Customers may not be able to inspect every platform layer or underlying log.

Clawsolator / XMAN approach: Compare deployability, custody, audit export, contract terms and self-hosted options.

Verification boundary: Lack of visibility is not proof that a provider is insecure.

Evidence: Gumloop Trust Center ↗ · How claims are classified

RISK CLASS 15 / False compliance and certification

False compliance and certification

Failure mode: Badges and security language can imply an assessment that was never performed.

Clawsolator / XMAN approach: Publish test methods, scope and limitations; label OSSTMM inspiration accurately.

Verification boundary: Clawsolator does not claim independent certification or OSSTMM compliance.

Evidence: ISECOM OSSTMM methodology ↗ · How claims are classified

Illustrative conversation — not an actual agent execution log

Meet the Lobster Security Desk

SysOp Clawsolator lobster avatar
SysOp

“If the agent says it needs sudo, does that mean it gets sudo?”

Security Officer Clawsolator lobster avatar
Security Officer

“No. The model can request a change. The operator-approved daemon policy decides whether an exceptional host action is allowed.”

Master Clawsolator lobster avatar
Master

“And the receipt? We want the request, reason, scope, and execution outcome—not just the model saying ‘done.’”

Current controls versus future ambitions

Implemented in the inspected XMAN source

Manager-mediated operations, request/approval handling, resource and agent-configuration workflow, audit/log scaffolding. Validate installation before public claims about enforcement.

Upstream runtime controls

OpenClaw sandboxes and Podman; OpenShell external policy and network restrictions; provider IAM and host hardening.

Roadmap, not shipped

Security Officer plugin audits, runtime behavior detection, automatic incident containment, and expanded OpenShell integration.

Read the roadmap and its verification gates →

Strong claims deserve strong evidence.

Inspect the controls, verify the boundaries, and ask what happens after a failure—not merely before it.

Explore the security lattice →Read the architecture →

CANONICAL TRUST BOUNDARY

Agents request. XMAN decides. Operators authorize exceptional host work.

Each registered agent uses the authenticated requester API through the trusted Gateway bridge. The XMAN management daemon governs policy, resource admission and durable outcomes through its private Unix-socket interface. The operator API is separate; the master coordinates but does not become the operator. The future SysOp and Security Officer daemons would be separate, permissioned services, not current Alpha functionality.

Compare the right categories before you buy

OpenClaw, OpenShell, Hermes, ZeroClaw, Gumloop, Aider, Continue and cloud providers solve overlapping but different problems. Find what each actually provides, what you must still configure, and what XMAN adds.

NEW RESEARCH / DOCUMENTED OPERATOR FRUSTRATIONS

What users asked the systems to do

Firsthand issue reports, exact failure symptoms, the proposed engineering response and explicit verification limits.

We fixed automatic containerized updates. →The QR code worked. The dashboard disappeared. →HTTP works. The agent still cannot connect. →“Container started.” So where did it go? →A thousand popups are not a security philosophy. →One runaway worker should not own the whole house. →There’s disorder across the harbor. →