Protect one agent from repository to runtime.

1. Create a project and issue a runtime key

Open the control plane, choose the environment, publish the default policy and create a one-time API key. Production, staging and test projects start in enforce mode; development starts in monitor mode.

2. Add a Guard call at the trust boundary

const decision = await fetch('https://agentrisklayer.com/v1/guard', {
  method: 'POST',
  headers: {
    authorization: `Bearer ${process.env.AGENTRISK_API_KEY}`,
    'content-type': 'application/json'
  },
  body: JSON.stringify({
    request_id: crypto.randomUUID(),
    input: untrustedInput,
    tool_call: proposedToolCall,
    metadata: { application: 'support-agent' }
  })
}).then(response => response.json());

if (decision.decision !== 'allow') throw new Error('Security policy denied the action');

3. Record the asset baseline

Paste or submit the agent manifest in the Inventory section. A later snapshot that adds public, privileged or production exposure automatically changes the deployment gate to review required.

4. Scan the repository locally

curl -fsSLO https://agentrisklayer.com/downloads/agent-risk-inspector.mjs
node agent-risk-inspector.mjs scan . --authorised \
  --environment test \
  --out agentrisk-inspection.json \
  --sarif agentrisk-results.sarif \
  --fail-on high

5. Gate pull requests

Use the supplied workflow to scan changes, upload SARIF and fail the build on active high or critical findings.

Evidence and privacy

• Hosted runtime screening stores no raw prompt, response or tool arguments.
• Local inspection excludes source content and matched secret values from evidence.
• Project keys are one-time secrets and only their hashes are stored.
• Every policy, key, inventory and remediation change is audit logged.
• A passing result is engineering evidence—not an independent penetration test or certification.