70% of AI agent deployments ship with unrestricted access to your API keys, filesystem, and local network. Most operators never find out until something goes wrong.
GRITS Audit scans your agent's live configuration (OpenClaw and NVIDIA NemoClaw), scores your security posture across 5 layers, and patches what it can automatically -- with a full backup and one-command rollback before touching anything.
git clone https://github.com/X-Scale-AI/grits-agent-scanner.git
cd grits-agent-scanner
chmod +x grits-*
./grits-auditNo dependencies. Python 3 stdlib only.
Record a 10-second
./grits-auditGIF and drop it here.
Native tools like openclaw security audit act as smoke detectors -- they tell you there is a fire, but leave the high-risk manual patching to you. GRITS is the fire extinguisher. GRITS actively scans, scores, and remediates vulnerabilities across the five critical trust boundaries:
- Layer 1: Network (Blast Radius): Unrestricted LAN exposure. We detect and help restrict agents from port-scanning or attacking your private
192.168.x.xhome or VPC subnets. - Layer 2: Operator (Identity): Missing trust boundaries. We enforce cryptographic Telegram ID allowlists so unauthorized users can't hijack your agent.
- Layer 3: Application (Plugin Sandbox): Unauthorized tool access. We lock down permissive tool policies to strict, explicit allowlists.
- Layer 4: OS & Secrets (Credential Vault): Credential leakage. We flag plaintext API keys sitting in
openclaw.jsonand enforce system-level secret injection. - Layer 5: Financial (Cloud Cost Bleed): Unthrottled API usage. We patch bloated contexts and re-route expensive background heartbeats to free, local SLMs.
Run ./grits-audit to see exactly what you're exposing and what can be patched automatically right now.
Patching a live AI agent is risky. One malformed JSON bracket can crash the gateway. ./grits-audit eliminates that risk using a stateful three-step workflow:
- Scans: Evaluates your live config and scores your zero-trust posture.
- Safely Patches: Takes a complete snapshot of your state before mutating anything. Applies the auto-fixes with your confirmation. (If anything breaks,
grits-audit --rollbackrestores your exact previous state instantly). - Re-scans: Verifies the applied patches and generates your updated compliance score.
Currently supports OpenClaw and NVIDIA NemoClaw.
A secure baseline is useless if a developer accidentally opens a port, tweaks a plugin policy, or commits a plaintext API key next week. Configuration drift is the #1 cause of compromised AI agents.
GRITS isn't just a local developer tool; it is an automated gatekeeper designed to run headlessly in your CI/CD pipelines (GitHub Actions, GitLab CI, etc.).
- Block Insecure Deployments: Instantly fail the build and block pull requests if an agent's Zero-Trust boundaries are degraded.
- Zero-Touch Auditing: Generate machine-readable compliance reports (JSON, SARIF) on every single commit.
- Enforce the Standard: Guarantee that no agent hits production unless it strictly adheres to your 5-Layer Security Model.
To configure GRITS for headless pipeline execution and advanced reporting, run:
./grits-audit --yes --scan-only # headless audit, no prompts
./grits-agent-scanner --json # machine-readable output for pipelinesGRITS patches what can be scripted. Network segmentation, SIEM integration, zero-trust architecture, and compliance mapping require decisions specific to your infrastructure and your threat model.
If your scan output flags Advanced Configurations -- or if you are deploying agents in a production or regulated environment -- that is where X Scale AI comes in.
Book a free architecture review
See CONTRIBUTING.md. Issues and PRs welcome. Vulnerabilities: SECURITY.md.
Built by X Scale AI | Apache 2.0 | Free architecture review