
Every agent carries a copy of your AI key
To call a hosted AI model, an agent today carries the company's own provider key in its settings. Fifty agents means fifty copies, in laptops, scripts and pipelines that nobody inventories.
- Copy a settings file, and you copy the key. That one key unlocks, and pays for, the company's whole AI account.
- AI keys are among the fastest-growing leaks. GitGuardian counted 1,275,105 AI-service secrets leaked on public GitHub in 2025, up 81% in a year, and 24,008 unique secrets in the configuration files that connect AI tools to other systems.¹
- Leaked keys stay live. Of the secrets GitGuardian found valid in 2022, more than 64% still worked in January 2026.¹
- A leaked key widens what an agent can do. OWASP's Top 10 for Agentic Applications names identity and privilege abuse, where leaked credentials let agents operate far beyond their intended scope.²
- Stopping one agent breaks all of them. If every agent shares the key, cutting one off means changing the key everywhere.
The fix is not more copies, carefully hidden. It is no copies at all in the agents' hands.
Your key stays in your vault
With Argent Strategic Intelligence, the company's AI provider key lives in one place the company controls. By default, the agents that work through Argent never hold it.
- Your own vault. Argent connects to the vault you already run: HashiCorp Vault or OpenBao, AWS Secrets Manager, CyberArk Conjur and Infisical. Argent records a reference to the key and resolves it on use; it never stores a copy of the value. See every integration.
- Or Argent's built-in vault. A company without its own vault uses the vault Argent ships with. The key is pasted once, written to the vault and never shown again, and no key value is stored in Argent's database.
- No new tool to adopt. Your security team keeps managing keys where it already does.
- Each tenant stays in its own lane. A tenant admin can only point at secrets inside its own tenant's area of the vault. A request for another tenant's path is refused and signed.
- Adding a key is guarded. Adding a secret from the console takes a passkey confirmation, and a customer vault stays read-only to Argent until someone deliberately turns on writing.
Instead of the provider's key, each agent gets its own Argent key: shown once when the agent is registered, with only a hash kept by Argent. It tells Argent which agent is calling, and nothing more.
How a request gets its key, step by step
The agent points its AI client at the Argent gateway and sends its Argent key. Everything else happens at the gateway, on every AI model call.
- The agent calls the model through Argent. Its Claude Code setup reads "Send every call to your Argent gateway, not to Anthropic.", with its Argent key in the place its AI client already puts a key.
- Argent checks who is calling. The Argent key must belong to an active agent. The Argent key itself stops at the gateway: the AI provider never receives it.
- Argent checks the agent may use the key. A credential can require a permission; an agent without it is refused. Only agents allowed to use a credential get it.
- Argent fetches the key from the vault. No credential on file, a vault that cannot be reached, or an empty secret means a refusal with a named reason. A customer vault never falls back to another vault, and a failed vault never falls back to the agent's own key.
- Argent removes any provider key the agent sent. Every header that can carry a provider key is stripped, so a client cannot smuggle its own key past the gateway. A key pasted into the web address is refused and scrubbed.
- Argent attaches the vault's key on the way out. For AWS Bedrock Argent signs the request itself, and for Azure OpenAI it gets the access token itself, so agents hold no cloud keys either.
- Argent signs a record. It names the agent, the provider and the vault reference. It never contains the key.
This covers the AI model calls of every agent that works through Argent: AI coding agents, customer-facing assistants and enterprise copilots.
Stop one agent, not all of them
Because every agent holds its own Argent key, one agent that goes wrong can be stopped on its own. The company's provider key is not touched.
| Action | What happens |
|---|---|
| Kill | The admin types the agent's name to confirm, and the console reports: "Its key no longer authenticates." Every other agent carries on. |
| Quarantine | The agent is refused until an admin releases it, whatever budget it has left. |
| Pause | A reversible soft stop, with a stated reason, resumed later. |
| Rotate the Argent key | The new key is shown once; the old key stops on the agent's next request. |
| Bind to a network | An admin lists the networks an agent's key may be used from; requests from anywhere else are refused. |
When the provider key itself must change, the company changes it in its own vault, or rotates it from the console, which writes the new value and then lets the admin keep it or roll it back. Revoking a provider credential takes the word REVOKE typed in full: Argent drops its reference, and the value stays in the vault.
Every one of these actions taken in the Argent console is a signed record. Pairs with identity tethering, which ties every agent to a named employee, and owner offboarding, which stops a leaver's agents.
What the operator sees
The Credential Broker page in the Argent console is one place for every vault, secret and use of a secret.
| Where | What it shows |
|---|---|
| Credential Broker header | "Key injection On" and the passthrough guard setting, so the posture is plain at a glance. |
| Vaults and secrets | Every connected vault and every reference, by kind and use, with totals. If any part cannot be read, the page says so rather than showing a false zero. |
| Uses and last used | Which agents used a secret, their human owners, and the last five calls. |
| Secret detail | Test that a reference resolves, repoint it, see its impact, rotate it, or revoke it. |
| Agent page | The agent's status and its controls: Quarantine, Pause, Cap, Kill, with every change on the agent's audit tab. |
| Signed record | Agent, provider, vault reference and header names, never the key. A refusal carries its reason code, such as credbroker.no_cred, credbroker.unauthorized or credbroker.vault_unavailable. |
When an auditor asks which agents used the company's AI key last month, the record answers. Go deeper with session replay and one-click audit evidence.
Where it fits, and how you start
Keyless agents serve the goals that come up first when agents reach production.
- Goals: it is central to agent identity and ownership, and it supports threat defense, audit and accountability, compliance evidence and AI visibility.
- Industries: banks and financial services, insurance, healthcare, the public sector and retail can keep the key to their AI account in their own vault. All industries.
- Regulated and sovereign AI: run Argent and its vault inside your own network. See AI in regulated and air-gapped environments, sovereign AI and deployment options.
- Control what agents reach next with agent reach control and model control.
- Watch. In Watch mode an agent's own key passes untouched, and Argent records what it would have stripped or refused. See watch before you enforce.
- Connect. Point Argent at your vault, or use its built-in vault, and add the provider key once.
- Register. Give each agent its own Argent key, point it at the gateway, and retire the old copies.
- Enforce and prove. Switch on Enforce. Every request's record shows which agent used which key, and why.