The pitch for AI agents is autonomy: systems that plan, call tools and complete work without a human in the loop. The security consequence is less discussed. An agent needs credentials to act, and most enterprises have no mechanism that issues them per action, scopes them to a task, or revokes them when the task ends. The result is a growing population of long-lived API keys and service accounts held by software that no one has inventoried, governing access that no one has reviewed.

Why a password vault is not enough

Credential vaults solved a human problem: too many administrators, too many shared passwords, no record of who used what. They were designed around a session that begins when a person authenticates and ends when they log off. An AI agent does not work that way. It may run continuously, spawn parallel tool calls, delegate to other agents, and operate across cloud, SaaS and internal systems in a single workflow. Vaulting the agent’s static credential simply moves the risk: the secret still exists, still persists, and is still usable by anyone who extracts it from the agent’s runtime.

The deeper issue is authorisation granularity. A password vault can say yes or no to a checkout. An agentic workflow needs an answer to a harder question: may this agent, acting on behalf of this user, for this task, invoke this specific API with these parameters, right now? That is a policy decision at the level of the action, not the credential.

Three controls that actually apply

Ephemeral, task-scoped credentials. Modern privileged access management extends into secrets brokerage for non-human identities: short-lived tokens minted at the moment of need, bound to a workload identity, and expired within minutes. The agent never holds a durable secret, so exfiltration yields nothing reusable. Cloud workload identity federation and dynamic secrets engines in platforms such as HashiCorp Vault are the practical building blocks, with the PAM layer providing discovery, approval and audit across them.

Runtime authorisation on the action. Vendors including Delinea, CyberArk’s Idira platform and Keeper have shipped policy engines that inspect what an agent is attempting before it executes, rather than authenticating it once at the start. The design pattern is familiar from database firewalls and web application firewalls: parse the intent, match it to policy, allow, block or escalate. Applied to agentic access, it gives security teams a place to express rules such as “agents may read customer records but never export them” without rewriting the agent.

Human approval for consequential actions. Not every agent action should be automated. The emerging pattern, visible in recent Yubico and Delinea work, links high-impact agent operations to an out-of-band human approval, ideally bound to a hardware authenticator so that the approver’s identity is strongly established. This preserves autonomy for routine work while keeping a human accountable for irreversible steps.

What to inventory first

Before policy design, organisations need an accurate map of agent identities. That means extending discovery beyond human directories to service principals, OAuth grants, API keys and workload identities, and attributing each to an owner, a purpose and a data scope. Most programmes discover that a substantial share of agentic access is unowned. That inventory, not the policy engine, is usually the limiting factor in reducing risk, and it is the same discipline privileged account security has demanded of human administrators for two decades.