
Argent with Google Cloud
Run Argent in your own Google Cloud project, send Google Gemini calls through it, and let Argent read your project to show whether Google itself refuses AI calls that go around the gateway.
Argent across Google Cloud
Argent Strategic Intelligence is the gateway between your AI agents and the AI models they use. It meets Google Cloud in five places.
| Where | What Argent does |
|---|---|
| Runs in your project | One Helm chart on Google GKE. The gateway's edge handles TLS itself, and the secrets store unlocks from Google Cloud KMS through the cluster's workload identity. |
| Governs Google Gemini | The gateway holds the one real Gemini key and attaches it on the way out; agents hold only their own Argent key. See Google Gemini. |
| Agents with no keys | A workload can send a service-account ID token from the metadata server. Argent checks the token and maps it to an agent with a human owner. |
| Evidence on Google storage | The signed evidence archive can live in Google Cloud Storage, signed in with the pod's workload identity, so no storage key is kept anywhere. |
| Checks the cloud | A read-only coverage scan reads your project and reports whether Google itself blocks AI calls that skip Argent. The rest of this page is about that scan. |
Why the cloud matters. A gateway governs what goes through it. Google Cloud can make it the only way. Google describes VPC Service Controls as letting you "define security policies that prevent access" to services from outside a perimeter (Google Cloud). Organization policies can switch off the Gemini Developer API in projects where Argent is not the caller, and limit which models may be used.
What the scan checks in Google Cloud
Every check is read-only. Each one ends in pass, fail or unknown, and the evidence says why.
| Check | Pass means |
|---|---|
| An enforced perimeter guards the AI platform API | The project sits inside an enforced VPC Service Controls perimeter whose restricted services include aiplatform.googleapis.com. A perimeter that exists only in dry run is a fail, named as such in the evidence. |
| An approved-models list is set | Google's approved-models organization policy has a non-empty allow list, or denies all. The allowed entries are shown in the evidence. |
| The Gemini Developer API is off | The effective service-usage organization policy denies generativelanguage.googleapis.com. |
| API keys stay away from the AI APIs | Every API key in the project is restricted to services other than the two AI APIs, or the project has no keys. Each key gets its own finding. Argent's own Gemini key shows up here too, so scan the projects your agents run in, not the project that holds Argent's key. |
Dry run is a fail, on purpose. Google's dry-run mode lets you "monitor requests to protected services without preventing access" (Google Cloud). That is the right first step, and the scan names it, but nothing is refused until the perimeter is enforced. A policy rule with a condition, or a value given as a group, reports unknown.
How a scan runs, and what you see
Argent reads the account, judges it, and signs the result.
- Argent signs in read-only. It fetches the scan credential from the Credential Broker by its reference.
- Argent runs every check. On its schedule, every 24 hours by default, or when an admin presses Scan now.
- Argent gives a verdict. Bypass blocked when every check that closes a path passes. Bypass possible when one fails. Unknown when none fails but one could not be read. All four checks count toward the verdict.
- Argent signs the result. Each scan writes a signed record, and every check whose result changed since the last scan writes another.
- Argent raises an alert on a drop. When an account falls from blocked to possible, an alert opens with the failing checks, and an admin acknowledges it. A return to blocked is signed too.
The scanner talks only to Google. It refuses a key whose service account email does not match the one registered, signs in only at Google's own token endpoint, and talks only to the organization policy, access context manager and API keys services.
What you enter in the Argent console
An admin adds the account under Integrations > Discovery > Cloud coverage, with Add account. The form asks for:
| Field | What it is |
|---|---|
| Cloud | Google Cloud. |
| Name | Optional. Shown instead of the project id. |
| Project id | The project to scan. |
| Organization id | Optional. Without it the perimeter check reports unknown. |
| Scan service account email | The service account the scanner signs in as. |
| Credential reference | Where the service account's JSON key is kept in the Credential Broker, never the key itself. |
| Scan every (hours) / Scan on schedule | How often the scan runs on its own. The default is every 24 hours. |
- See every account at a glance. The page counts accounts by verdict, shows each account's last scan, lists open alerts and explains what each check reads and what a pass proves.
- Who can do what. SuperAdmins and tenant admins add accounts and run scans. Auditors can read every result.
- One switch, with a lock. Scanning has one switch per tenant: on, off or locked. Off skips the schedule and refuses Scan now, and deletes nothing. A higher level can lock the switch, and only that level can change or unlock it. Every change is signed, and so is every refused one.
What you set up on the Google Cloud side
One read-only identity, once. Nothing on this list lets Argent change your cloud.
- A service account for the scanner with three read-only roles:
roles/orgpolicy.policyViewer(project or organization),roles/serviceusage.apiKeysViewer(project) androles/accesscontextmanager.policyReaderat the organization, for perimeters. - Three APIs on in its home project:
orgpolicy.googleapis.com,apikeys.googleapis.comandaccesscontextmanager.googleapis.com. If one is off, Google refuses the read and the check reports unknown with the reason. The scanner never turns anything on. - Its JSON key, saved as one reference in the Credential Broker. Give Argent the reference and delete the local copy.
- The levers themselves are yours. Your Google Cloud admins set the perimeter and the organization policies, starting with one test project, and with a dry run wherever Google offers one (the perimeter and the service-usage policy).
Where it fits, and how you start
The scan is how a security team proves the gateway is the only way to the models, not just the easiest one.
- Capabilities: shadow AI discovery, model control, keyless agents, watch before you enforce and one-click audit evidence. All capabilities.
- Goals: see all AI in your company, model control and routing, threat defense, compliance evidence and sovereign AI. All goals.
- Use cases: AI coding agents, customer-facing assistants, enterprise copilots, agent tool access and regulated and air-gapped AI.
- Related integrations: Google Gemini, Kubernetes and Helm, and the scans for Amazon Web Services and Microsoft Azure.
- Where Argent runs: air-gapped on premises, private cloud or your public cloud VPC. Compare the options.
- Industries: financial services, insurance, healthcare, the public sector and retail. All industries.
- Install. Put the Helm chart in your GKE cluster with the Google Cloud profile, and point Gemini callers at the gateway.
- Watch. In Watch mode, Argent records what it would have stopped and changes nothing.
- Scan. Add the account under Integrations > Discovery > Cloud coverage and press Scan now. Each failing check is named, with its evidence.
- Enforce and prove. Switch on Enforce, have your admins pull the levers, and scan again until the account reads Bypass blocked. The signed records show when it got there.
Want this running on your own use cases?
Book a demo