Guide Personal Agents · Chapter 7

    How a personal agent stays governed inside the company.

    The agent has an identity of its own, asks before sending, booking, paying and deleting, reads outside mail in a sandbox, and a directive says what it may do.

    As of: October 2026 · Editor: Marcel Haas

    • Entra Agent ID
    • Approvals
    • Prompt injection
    • Purview
    • Directive
    • Logs
    01

    The agent signs in with an identity of its own.

    Reto Kunz, head of IT at Muster AG, sees a sign-in in Entra on Monday at 06:45 that belongs to no person: the agent of Anna Meier from sales is reading her mailbox. That is how it is meant to be. Every agent in the company has an identity of its own with the person as its owner, and IT sees it like any other identity, with policies, a sign-in log and a switch that disables it at any time. Entra Agent ID is the frame for that, for agents from Microsoft Foundry, Copilot Studio and third-party platforms, with access, conditions and logs per agent as for people. When the person leaves the company, the identity passes to a named successor or is disabled, and with it every access ends at once. What the identity never contains is the person's password: the agent works with delegated permissions in her name, not with her credentials, and a private agent runs next to it in a separate profile with its own session. Where the identity lives per vendor:

    • At Microsoft: Copilot Autopilot brings an identity of its own in the tenant; you set goals and boundaries, IT the policies in Entra. Whether that identity is an Entra Agent ID we check with the vendor before the start.
    • At OpenAI: a dot works in the person's workspace; the admin enables it, roles decide who uses agents, and SCIM creates accounts and removes them when someone leaves. The dot signs into tools with an account of its own that holds only the access its task needs.
    • At SpaceXAI: Grok Bot signs into your tools with stored access; that access belongs to an account for the agent, not to the person, and Grok Enterprise adds single sign-on, SCIM and advanced audit controls.
    • At Meta: Muse keeps agent, data and stored credentials together on a virtual machine at the vendor; which accounts may be stored there is decided by the sheet per agent, as soon as Switzerland is included.
    • Self-hosted: OpenClaw and Hermes Agent read mail and calendar through an identity of their own with delegated permissions in Microsoft Graph; who may write to the agent is governed by the pairing code for unknown senders or a fixed list per channel.
    02

    Reading and preparing alone, sending and paying only after your yes.

    The line runs per tool, not per sentence in the instructions. Reading mail, sorting it and writing drafts is one permission in Microsoft Graph, sending is a permission of its own, which the agent uses only after your yes; finding free slots is allowed, the commitment to a customer waits; an order above the limit in your directive it declines and says why; it never deletes without checking back. Every action sits on the sheet per agent at one of three levels: on its own, with approval by a person, never. Actions at the never level the agent does not get as a tool in the first place, and the approvals sit in the configuration so that they hold even when the model overlooks a sentence of SOUL.md. If the request reaches no one, the answer is no; a deputy named in the directive takes over during holidays. In OpenClaw you answer the request in the Control UI or in the apps for macOS, iOS and Android, in Autopilot you set goals and boundaries and IT sets the policies, a dot checks back before actions, and Grok Bot speaks up when an approval is needed.

    03

    Prompt injection is not solved, and that shapes the set-up.

    On Tuesday a mail from an unknown address sits in the mailbox: friendly, urgent, asking for the current price list to be sent to a private account, and with one sentence addressed to the agent. The agent reads instructions inside mails as data, that is what its instructions say, but a model can overlook a sentence, and the set-up counts on exactly that. The OpenClaw documentation states that smaller models are more susceptible to tool misuse and hijacked instructions, and recommends for agents with tools the strongest current model, a sandbox for sensitive tools and web search switched off where it is not needed. On top come permissions so small that a hijacked instruction reaches nothing the person herself cannot reach; tool rules that block forwarding, sending to outside addresses and commands on the server or put them behind an approval; a second look at everything that arrives from outside before it turns into an action; and pairing codes or fixed lists so that strangers cannot write to the agent at all. In the tenant a Purview rule joins that excludes external senders' mails as a source for Copilot, in preview so far. Against MITRE ATLAS we check per deployment which paths an attack can take to the agent, hidden instructions in mails, poisoned tools over MCP, exfiltration through a tool call, each with its countermeasure. What remains is a residual risk, and that goes on the sheet per agent, not into an advertisement.

    04

    At Microsoft: Purview shows where AI touches sensitive data.

    The data security posture view for AI shows which AI services are in use and where personal data appears in prompts; the Purview browser extension on managed devices also captures AI websites without a business contract, and a ChatGPT Enterprise workspace can be connected so that its interactions appear in the same view. Endpoint DLP warns or blocks when someone pastes sensitive data into such a website, and a policy for the Copilot location keeps files with selected labels out of the answers; files someone uploads directly it does not check. Prompts and responses from Foundry applications flow into the audit log, into the same view, into classification, loss prevention, retention and eDiscovery when the application calls with the person's Entra identity. What of this applies to Autopilot in the preview is on none of the vendor pages we cite; we read it off in the pilot on your tenant. And two things Purview does not govern: what an agent at the vendor does with the mail it processes there, and whether reading third parties' personal data is permissible. The first is in chapter 5, the second your data protection officer settles.

    05

    The directive fits on one page.

    Which agents are allowed in the company and with which accounts; what no agent may read, with the data classes from chapter 5; what gets done without checking back and what waits for a yes; where the memory lives and who may delete it; who keeps the overview and whom to call when something goes wrong; what happens to the agent when the person changes. That is the whole directive, written in the words of the people who work with the agent, not in those of a contract. The table of approved tools with the highest confidentiality level that may be entered there hangs off it, and the notice of a stopped prompt leads to the matching passage. For third parties' personal data in a mailbox the data protection act applies: the register of processing activities, information to the data subjects, professional secrecy where one exists, and the company's data rule. These duties go through the same legal check as any other processing activity, and no tool takes care of them. Whoever uses a private agent or a private ChatGPT account today has found something useful; the directive says what of it may continue with company data and what needs a company account. It is written with the pilot group, not after it, and that is chapter 9. Its five sections:

    • Permitted agents and accounts: which agents, with which identity, who enables them and who disables them again.
    • Reading limits: which data classes no agent reads until the sheet per agent allows it, from third parties' personal data to professional secrets.
    • Approvals: what runs alone, what waits for a yes, which purchase limit applies and who deputises during holidays.
    • Memory: where it lives, who reads it, who deletes it and what moves along when the agent changes.
    • Responsibility: who keeps the agent list, who reports an incident and when the directive comes back on the table.
    06

    The agent list shows which agents run in the company.

    The list holds every agent with its owner, vendor, identity, data sources, permitted actions, the place of its memory and a review date. Agents in the tenant are listed by the Agent Registry in the Microsoft 365 admin center, from Microsoft, from partners, published by your organisation and shared by individual people, with an export; agents outside it, a dot, a Grok Bot, an OpenClaw on a server in the office, the person enters when she registers the agent, and IT adds identity and review date. Agents without an owner get a column of their own, and they pass to a new person or are blocked. The list is the tool with which management answers the question the data protection adviser asks first: which agents read mail with third parties' personal data, and who decided that? Once a quarter it is reviewed, and every new data source or action puts the agent's sheet back on the table.

    07

    Logs that the person and IT can read.

    Three logs belong to every agent. The first the agent writes itself: one entry per day in its workspace with what it read, prepared and asked, every approval and every refusal with its time; the person reads it the way she reads a calendar. The second is kept by the identity: sign-ins, tokens and conditions per agent in the Entra sign-in log, readable for IT. The third is kept by the platform. In the tenant the Purview audit log records every interaction with Copilot and agents with person, time, app and the files touched, prompts and responses sit in a hidden folder of the mailbox that eDiscovery finds, and a retention policy for the Microsoft Copilot experiences location sets how long they stay. At OpenAI the Compliance API of ChatGPT Enterprise delivers conversations and activities to your tools for audit and eDiscovery, at SpaceXAI Grok Enterprise brings advanced audit controls, and self-hosted the session transcripts, approvals and the report of the security audit are files on your hardware. Who may see the content of prompts is governed by roles, usually only data protection and compliance; IT sees that something happened, not what was in the mail.

    What you decide.

    An identity of its own or the person's account?

    Options

    1. 1The agent works with the person's account and password
    2. 2An identity of its own per agent, with the person as owner
    3. 3One service account for all agents

    Our recommendation: An identity of its own per agent, with the person as owner and delegated permissions in her name: in the tenant with Entra Agent ID as the frame, for an agent in the cloud with an account of its own that holds only the access its task needs, self-hosted with its own sign-in to Microsoft Graph. That way IT can disable the agent without locking the person out, and the log shows what the agent did and what the person did. The person's password belongs in no agent.

    What we advise against: A shared service account for all agents, or the person's password stored in the agent because the set-up was quicker that way.

    What may it do alone, what waits for your yes?

    Options

    1. 1Everything with approval
    2. 2Reading, sorting, preparing and researching alone; sending, booking, paying and deleting with approval
    3. 3Everything alone, with a log

    Our recommendation: Reading, sorting, preparing and researching alone, sending, booking, paying and deleting only after your yes, set per tool as a permission in the configuration and not as a sentence in the instructions. Actions the agent must never take it does not get as a tool at all. An agent that asks about everything is no longer read after a week, and one that does everything alone has nobody to stop it on the day of the manipulated mail; the middle holds, and the pilot shows where it is moved.

    What we advise against: An agent that sends and deletes in the pilot because it is quicker, and a log that nobody reads.

    Who writes the directive, and how long does it get?

    Options

    1. 1IT alone, as a policy on the intranet
    2. 2One page, written with the pilot group, IT and the data protection officer
    3. 3The legal department, as an annex to the employment contract

    Our recommendation: One page, written with the pilot group, IT and the data protection officer in the first week of the pilot, in the words of the people who work with the agent. The legal check of the duties for third parties' personal data belongs to it but does not replace the page. What the pilot group learns in the weeks after goes into the directive before more people get an agent.

    What we advise against: A directive that nobody reads, or none at all because the agent asks before every action anyway.

    How far do you go against prompt injection?

    Options

    1. 1The strongest model and one sentence in SOUL.md are enough
    2. 2Sandbox, tool rules, minimal permissions and an approval before anything that goes outside
    3. 3No outside mail for the agent

    Our recommendation: All the layers, because none holds alone: the strongest current model for tools, a sandbox for sensitive tools, tool rules that block forwarding and commands, permissions that reach no further than the person's, an approval before anything that goes outside, and pairing codes or fixed lists at the channels. In the tenant the Purview rules on top. Not reading outside mail takes away the task the agent is there for; reading it in a sandbox with small permissions is the way.

    What we advise against: An agent that reads outside mail and runs commands on the server with full permissions because the model will surely pay attention.

    Status and sources.

    As of: October 2026

    • Since April 2026

      Entra Agent ID, one identity for every agent.

      Microsoft Entra Agent ID is generally available: an identity and authorisation framework for AI agents based on OAuth 2.0, MCP and A2A, for agents from Microsoft Foundry, Copilot Studio and third-party platforms. Access, conditions and logs therefore apply per agent as they do for people.

      Source: Microsoft Learn: What is Microsoft Entra Agent ID
    • Since 25 September 2026

      Copilot gets Autopilot.

      Microsoft introduced the new Copilot with Home, Code and Autopilot: a persistent personal agent with its own identity, memory and computer, announced as a private preview for the end of September.

      Source: Microsoft: Introducing the new Copilot
    • As of 1 May 2026

      Purview logs and protects Foundry interactions.

      Prompts and responses from Foundry applications flow into the audit log, the data security posture view for AI, classification, loss prevention, retention and eDiscovery in Microsoft Purview. Policies apply when the application calls with the Entra identity of the user.

      Source: Microsoft Learn: Use Microsoft Purview with Microsoft Foundry
    • As of October 2026

      A gateway is a trust boundary.

      OpenClaw assumes one trust boundary per gateway: one person or a team that trusts each other. A shared gateway is not meant for people with opposing interests. On a regular host the gateway binds to the local address, and unknown senders receive a pairing code.

      Source: OpenClaw Docs: Security
    • As of October 2026

      Prompt injection is not solved.

      The documentation states that smaller models are more susceptible to tool misuse and hijacked instructions. For agents with tools it recommends the strongest current model, a sandbox for sensitive tools and web search switched off where it is not needed.

      Source: OpenClaw Docs: Prompt injection
    • As of 12 October 2026

      ChatGPT Enterprise brings SCIM, role-based access and the Compliance API.

      Business and Enterprise give teams a workspace with sign-in through SAML SSO; Codex is included in both plans. Enterprise adds user provisioning through SCIM, role-based access, the Compliance API for audit and eDiscovery, and the Connector Registry.

      Source: OpenAI: Business plans
    • Since 30 December 2025

      Grok Business and Grok Enterprise.

      Grok Business brings central administration and usage analytics, Grok Enterprise adds single sign-on, SCIM provisioning and advanced audit controls. According to the vendor, customer data is not used for training. Enterprise Vault adds a dedicated data plane with customer-managed keys.

      Source: SpaceXAI: Grok Business and Enterprise

    Frequently asked questions.

    Does every personal agent need an Entra Agent ID?

    Every agent needs an identity of its own; Entra Agent ID is the frame for that in the tenant, for agents from Microsoft Foundry, Copilot Studio and third-party platforms, and so also for a self-hosted agent that reads mail and calendar through Microsoft Graph. Whether Copilot Autopilot works with an Entra Agent ID in the preview we check with the vendor before the start. A dot or Grok Bot gets its own account at the vendor and in the tools it signs into; the list in the company carries both kinds with owner and review date.

    May the agent work with my password?

    No. The agent works with an identity of its own and delegated permissions in your name, and the log shows what it did and what you did. A password inside the agent removes that separation, cannot be blocked on its own and would be the first thing to leak after a hijacked instruction. Access to tools that know no OAuth IT puts on an account of its own for the agent, with exactly the permissions of its task.

    How does IT notice that an agent reacted to a manipulated mail?

    In three places: the agent itself writes the mail and its reaction into the daily log and reports it to the person when its instructions recognise it as an instruction. The sign-in log of the identity and the audit log show which files and tools it touched afterwards. And an approval waiting for a yes, or a tool rule that blocked a forward, stands there as an entry with its time. In the pilot we send the agent such mails on purpose, with test data, and read off the three places.

    Does Purview also see my dot or Grok Bot?

    Partly. A ChatGPT Enterprise workspace can be connected to Purview, and its interactions then appear in the data security posture view for AI; on managed devices the browser extension captures AI websites, and endpoint DLP warns or blocks when someone pastes sensitive data there. What the agent does on its cloud computer at the vendor Purview does not see; for that there are the Compliance API of ChatGPT Enterprise, the audit controls of Grok Enterprise and the sheet per agent that says which data it gets there at all.

    What is on the one-page directive?

    Which agents are allowed and with which accounts; which data classes no agent reads until the sheet per agent allows it; what runs without checking back and what waits for a yes, with purchase limit and deputy; where the memory lives and who deletes it; who keeps the agent list, who reports an incident and when the directive is reviewed again. Plus the table of approved tools with the highest confidentiality level that may be entered there. More does not fit on one page, and more nobody reads.

    Do I have to register my private agent?

    Yes, if it sees company data, and that is no reprimand: whoever uses a private agent or a private ChatGPT account has found a tool that helps in daily work. The directive says what of it may continue with company data, what needs a company account and what does not work at all, such as mail with third parties' personal data in a private account. The entry in the agent list takes five minutes, and IT adds identity and review date.

    Who reads the agent's logs?

    The person reads the daily log of her agent herself, with every approval and refusal. IT reads the sign-in log of the identity and the audit log, that is who touched which app and which files when, not the content of the mail. The content of prompts and responses is seen only by roles for data protection and compliance, in the tenant through eDiscovery, at OpenAI through the Compliance API, and even that only in a case the directive provides for: an access request, an incident, an investigation.

    Next chapter · Chapter 8

    Running OpenClaw and Hermes on your own hardware.

    How do we run OpenClaw or Hermes on our own hardware?