How are teams managing identities and access for AI agents?

Summary of public sourcesWritten by our research agent from 7 public sources. Not first-hand experience.

Agents often borrow a person's access. Practitioners say cloud IAM can give an agent narrow rights. Many SaaS tools offer only an all-or-nothing token, so agents end up on a broad human or project key. OWASP calls this excessive agency. Its advice is to give an agent only the tools, functions and permissions its task needs, act as the specific user with minimal OAuth scopes, and require human approval for high-impact actions.

Standards and vendors are adding structure. The Model Context Protocol's authorization spec requires tokens bound to the server they were issued for. It forbids passing a client's token through to other services. It also tells clients to request only the scopes they need. Microsoft's Entra Agent ID gives agents their own identity type, created from a template, with a human sponsor accountable for each agent. Admins can disable or revoke all agents of a kind at once. Practitioners also keep secrets away from the agent with a proxy that adds credentials, or run agents in a sandbox with narrow accounts.

Still open: how to get narrow access from tools that don't offer it, and where to put a hard check before an agent acts.

Sources

• Hacker News: Ask HN: How do you give AI agents access without over-permissioning?
• Hacker News: Ask HN: How do you provide your AI agents with access to credentials/secrets?
• Standard: LLM06:2025 Excessive Agency
• Standard: Authorization | Model Context Protocol
• Vendor docs: Overview of agent identities in Microsoft Entra
• Reddit: How are you giving AI agents access to AWS in production?
• Reddit: Heads up: Users can register agent identity blueprints by default