AI Agent Access Control (2026): The Tools, Agent Identity and What a CISO Should Require
Guide · · Published November 2026 · 9 min read
By Ashwin Gopinath, CEO
TL;DR
This page is for the CISO, CIO or Head of AI who has to decide what AI agents may see and do before they are rolled out. Access control for agents has four layers, and no single product covers all of them. Sentra, the organization memory for teams and AI agents, covers the knowledge layer: it applies each source tool's permissions to every fact an agent reads, and its agents wait for a person's approval before they send, post or change anything. Microsoft Entra Agent ID and Okta give each agent an identity. Auth0 and Amazon Bedrock AgentCore Identity handle an agent acting on behalf of a user. Cerbos, OpenFGA and Open Policy Agent decide whether one request is allowed inside the applications you build. Every vendor fact below was read on the vendor's own page on 18 November 2026.
What are the best tools for AI agent access control and permissions?
The right tool depends on the layer you are trying to control. The table groups nine by the question each one answers.
| Tool | Layer | What it controls | Source |
|---|---|---|---|
| Sentra | Permissions on company knowledge, and approval on actions | Which facts from Slack, email, meetings, documents, tickets and CRM an agent may read for a given person, and whether an action goes ahead | How permissions work |
| Microsoft Entra Agent ID | Agent identity | An identity account for each AI agent in Microsoft Entra ID, used for autonomous or delegated access | learn.microsoft.com |
| Okta | Agent identity and app-to-app access | Cross App Access, an OAuth extension Okta recommends for connecting to MCP servers in the enterprise | okta.com |
| Auth0 | Delegated access for agents you build | Agents acting on behalf of users with centralized authorization, where each action is auditable | auth0.com |
| Amazon Bedrock AgentCore Identity | Agent identity and credentials on AWS | Credentials that let agents reach AWS resources and third-party services on behalf of users, with audit trails | docs.aws.amazon.com |
| Cerbos | Authorization | Open-core policy decisions across applications, gateways, workloads and AI agents, with a log of every decision | cerbos.dev |
| OpenFGA | Authorization | Open-source relationship-based access control modeled on Google's Zanzibar paper | openfga.dev |
| Open Policy Agent | Authorization | An open-source, general-purpose policy engine for applications, proxies, Kubernetes and API gateways | openpolicyagent.org |
| MCP authorization specification | Standard | An open standard that binds each access token to the one MCP server it was issued for | modelcontextprotocol.io |
Identity providers and authorization engines answer who the agent is and whether a call is allowed. They do not know what was decided in last week's meeting or which sentence in a combined answer came from a private channel. That is the layer where company knowledge leaks, and it is the one a memory layer has to control.
What is AI agent identity, and how should an agent prove who it acts for?
Agent identity answers two questions: which agent is this, and on whose behalf is it acting. Microsoft's documentation describes the two modes plainly. With autonomous access, an agent uses access rights given directly to its own identity. With delegated access, an agent acts on behalf of a human user, using rights given to that user, and the user controls which rights are delegated (learn.microsoft.com).
- Prefer delegated access for anything that reads company knowledge. The agent then sees what the person could see and no more, and both identities appear in the log.
- Give every agent its own identity and a named owner. Microsoft records the user who creates a Copilot Studio agent as its sponsor, and says agent identities can be retired without leaving orphaned credentials behind.
- Bind tokens to one destination. The MCP authorization specification requires MCP servers to validate that an access token was issued specifically for them, which stops a token for one system being replayed against another.
- Never share one broad service account between agents. One leaked credential then exposes everything that account could reach.
In Sentra, an agent connected over MCP or the API acts as the person who connected it and inherits that person's access. Keys issued to third-party AI clients such as ChatGPT, Claude or Cursor can read memory but cannot use Sentra's write endpoints outside MCP.
How do you control what company information an AI agent is allowed to see?
Carry the permissions of each source system through to every fact the agent reads, and check them at the moment of the request. Copying documents into a separate index with its own access rules is where most leaks begin, because the copy drifts from the source.
- Inherit, do not re-create. If a person cannot open a channel, a document or a deal record in the source tool, an agent acting for them should not see facts drawn from it.
- Check at request time. Access changes when people move teams or leave. A check made only at ingestion is out of date the next day.
- Fail closed. If a permission check fails or times out, the answer is no access.
- Apply the rule per fact, after sources are combined. An answer that joins a public roadmap with a private salary thread has to drop the private part for a reader who cannot see it.
Sentra works this way. Every answer and action respects each source's existing permissions, through query-time checks and synced access lists, and a failed check counts as no access. Enterprise plans add role-based access and audit logs.
What should a CISO require before agents get access?
| Decision criterion | What good looks like | Question for the vendor |
|---|---|---|
| Permission inheritance | Source permissions apply to every fact, checked at request time | What happens when a permission check times out? |
| Agent identity | Each agent has its own identity, a named owner and delegated access | Can we see which person an agent acted for on each request? |
| Approval gates on actions | Sending, posting, paying and deleting wait for a named person | Which actions can run without approval, and can we change that list? |
| Audit | A record of who asked, what the agent read and what it did | Can we export the logs, and how long are they kept? |
| Deployment model | Cloud, a VPC or on-prem, matched to the data class | Can the knowledge an agent reads stay inside our own environment? |
| Data residency | Storage and inference regions named in the contract | Where does the model run when an agent reads our data? |
| Certifications | SOC 2 Type II and ISO 27001 covering the product you buy | Is this product inside the scope of the report? |
| Revocation | One switch disconnects an agent and its tokens | How fast does revocation take effect? |
Sentra holds SOC 2 Type II and ISO 27001, with cloud, isolated VPC, and fully air-gapped on-prem options. A named person approves before Sentra acts or sends anything. The longer checklist, including memory governance, is in AI agent permissions, and the regional questions are in AI data residency.
Which layer should you buy first?
- You are rolling out assistants and agents that read company knowledge from many tools: start with permissions on that knowledge, because it is where a leak reaches the most people. Sentra is built for this layer.
- You run Microsoft 365 and Copilot Studio: Microsoft Entra Agent ID is available to all Microsoft Entra customers and gives those agents identities.
- Your workforce identity is in Okta: look at Cross App Access for agent connections to MCP servers.
- Your engineers are building agents into your own product: add Auth0 or AgentCore Identity for delegated access, and Cerbos, OpenFGA or Open Policy Agent for the authorization rules.
FAQ
What is AI agent access control?
AI agent access control is the set of rules that decide what an AI agent may read and what it may do. It has four layers: an identity for the agent, authorization for each request, permissions on the company knowledge the agent reads, and human approval for consequential actions.
What are the best tools for AI agent access control and permissions?
By layer: Sentra for permissions on company knowledge and approval on agent actions, Microsoft Entra Agent ID and Okta for agent identity, Auth0 and Amazon Bedrock AgentCore Identity for delegated access in agents you build, and Cerbos, OpenFGA and Open Policy Agent for authorization rules.
What is AI agent identity?
AI agent identity is a distinct, managed identity for a software agent, separate from the people who use it. It lets a company see which agent made a request, on whose behalf, and retire the agent without leaving credentials behind.
Should an AI agent use the user's permissions or its own?
For reading company knowledge, the user's. Delegated access means the agent sees only what the person it acts for could see. An agent's own standing access suits narrow scheduled jobs, and it should be scoped to the least it needs.
How do we stop an AI agent from sending or changing things on its own?
Put an approval gate on every action with consequences outside the conversation. In Sentra, a named person approves before the agent acts or sends anything.
Does single sign-on solve AI agent access control?
No. Single sign-on proves who the person is. It does not decide which facts an agent may combine into an answer, or whether an action should go ahead. Those need permissions on the knowledge itself and approval on actions.
The decision rule
Buy identity so you know which agent did what for whom. Buy authorization for the software your own engineers build. And before any agent reads across Slack, email, meetings and the CRM, make sure the memory it reads from enforces the source permissions on every fact and holds actions for a person's approval.
To see how Sentra applies your existing permissions to what agents read, book a demo.