auditing-nemoclaw-policies

name	auditing-nemoclaw-policies
description	Reviews NemoClaw sandbox security policies for gaps, overly permissive rules, and best practice violations. Use when auditing policy security, before deploying policies to production, or during security reviews.

Auditing NemoClaw Policies

NemoClaw sandbox policies control filesystem access, network egress, process privileges, and enforcement behavior. The blueprint sets an initial policy at launch time; this skill audits runtime modifications to catch security regressions.

Audit Workflow

The audit follows a feedback loop: fetch, check, fix, re-validate, apply.

1. Fetch the current policy

openshell policy get <name>

This runs on the host and retrieves the active policy YAML for the named sandbox.

To inspect a previous revision:

openshell policy get <name> --rev N

2. Run the audit script

python scripts/audit_policy.py <policy.yaml>

The script reads the policy YAML and checks for common security issues. It produces a report with CRITICAL, WARNING, and INFO severity levels.

3. Review findings and fix issues

Address all CRITICAL findings before proceeding. WARNING findings should be reviewed and resolved where possible. INFO findings are informational and may be intentional.

4. Re-validate the updated policy

python scripts/audit_policy.py <policy.yaml>

Re-run the audit to confirm all CRITICAL issues are resolved and no new issues were introduced.

5. Apply the updated policy

openshell policy set <name> --policy <file> --wait

The --wait flag blocks until the sandbox confirms the policy is active. Use --timeout 120 for policies that take longer to propagate.

What the Audit Checks

The audit script inspects the following areas:

• Network policies: Wildcard or overly broad hostnames, missing TLS on port 443, shell binaries with network access, access: full grants, wildcard HTTP methods (method: '*'), missing binary restrictions, duplicate endpoints across policies
• Filesystem policies: Read-write access on sensitive paths (/etc, /usr, /lib, /proc)
• Process policies: Non-root sandbox user requirement
• Landlock: Missing landlock configuration
• Enforcement: Enforcement mode set to audit instead of enforce
• Empty policies: Empty network_policies (may be intentional for airgapped sandboxes)

Severity Levels

Level	Meaning	Exit Code
CRITICAL	Must be fixed before deployment	1
WARNING	Should be reviewed and addressed	0
INFO	Informational, may be intentional	0

The script exits with code 0 if no CRITICAL issues are found, and code 1 if any CRITICAL issue is present.

Detailed Reference

Security best practices checklist: See security-checklist.md