Blog · 18 November 2026
My AI harness signs in as me. That is the feature and the problem.
A practitioner note on connecting ChatGPT, Claude and Grok to CMDB, vulnerability management, threat intel and XDR: whose credentials, what happens when colleagues want in, and what you can actually log.
By Marcus Bengtsson, CTO at Sciber
I use ChatGPT, Claude and Grok the way many security people now do. I connect them to the CMDB, vulnerability management, external threat intel and XDR alerts, and I ask the model to help me solve problems and show me what I am looking at. I am still the only user on that harness.
It is useful. It is also an identity problem dressed up as a productivity win. The simplest picture I have is this: the harness borrows my name badge and my keys.
I recently wrote about agent computers: where the agent runs, what it can reach, and whether Conditional Access even sees the session. This note is the next step: what happens when the harness is pointed at the tools a security person lives in every day.
How I use it today
The pattern is simple. I open the chat, connect the tools I need for the question in front of me, and keep the loop short: pull context, look at the analysis, and check the output before anything leaves the chat.
I have not handed this to the team yet. That is deliberate. It is also my way of admitting I will not scale a setup I cannot explain or log. Once a second person shares the same connections, identity and logging stop being theoretical.
I am writing this as a practitioner asking peers how they run it, not as a vendor checklist.
Whose identity holds the token?
By default, the harness does not get an identity of its own. Most setups drift into that, because it is the easy path. That is no excuse for leaving it that way once other people want in.
On ChatGPT Business and Enterprise, after an admin enables an app, each user authorises their own account. ChatGPT then works inside that user's existing permissions. Custom MCP follows the same idea: users authenticate to each app before first use when OAuth is required. Admins can also set up administrator-managed or no-auth custom apps, so personal OAuth is not the only path.
Claude connectors inherit each person's permissions from the connected service. On Team and Enterprise, an Owner enables the connector for the organisation, and each person still authenticates individually unless you use Enterprise-managed auth. Claude also supports a static credential in request headers: the Owner pastes an API key once, org-wide. Their own docs say to treat that as the organisation's credential, not a person's, and to use OAuth when the tools must act as the individual.
Grok connector management has the admin add the connector first; members then connect their own accounts. On Grok Bot, the Bot has no identity of its own and acts as the signed-in member, with an exception for team-managed connectors that may use team or service-account credentials.
So the honest default is this. The model calls the CMDB and the XDR as you, or with a shared key someone pasted. MCP does not mean read-only. Write tools show up in vendor docs for security products. On Team and Enterprise plans, an admin usually has to enable the connector before you can use it at all.
That is fine for a careful loop with one user. It is a bad foundation for a shared SOC harness.
When colleagues want in
I see five options. None of them are free.
The first is that each person signs in with their own account. That is the closest match to each person's existing permissions. The weak point is telling "I clicked Isolate" from "the model called Isolate while I prompted," unless you add extra claims and a way to tie the events together. Cost rises with every seat.
This is the option I worry about most, because the agent inherits everything the person can do. If that person is an admin, the agent is an admin too, and every action it takes runs with those rights. I would not recommend it for anyone with high privileges. The fix is not exotic. Either trim the rights on that account to what the work actually needs, or give the harness a second account with only the access it should have. A neater answer would be for every access through MCP to get lower rights than the same person has in the application. That only works if the MCP server for that application supports it, so check before you count on it.
The second is one shared service account. Easy on day one. OWASP LLM06 (Excessive Agency) exists for a reason: one secret, and if it leaks, everything that account can touch is at risk, and you can no longer say who did what. I would not run a SOC on that.
The third is admin-managed connectors in the vendor workspace. ChatGPT, Claude and Grok all let admins gate what is available. Useful, and not sufficient on its own. You still need limits on actions, approval on writes, and a clear data processing agreement.
The fourth is a central internal MCP gateway. Tool traffic goes through a gateway that stamps a user id, an agent id and a correlation id, brokers short-lived tokens, and holds secrets in a vault. More build work. Best control when you wire a CMDB and an XDR to a language model.
The fifth is Microsoft Entra Agent ID, or an equivalent workload identity. Microsoft's agent identity is a specialised service principal from a blueprint. The agent identity has no credentials of its own; the blueprint acquires tokens. That is a real destination for a harness the organisation owns. It is not what ChatGPT, Claude or Grok Bot give you when a person connects a CMDB via their own sign-in. Check what is available in your tenant before you promise it. Some of the admin experience has still been in preview, and fuller agent security features sit on Microsoft Agent 365.
My working recommendation by stage is simple. Solo: personal sign-in only if your account is not a privileged one, otherwise a separate account with fewer rights; read-only access; no shared secrets. Team: organisation-managed connectors or a gateway, a separate identity per tool, and approval on writes. Never one shared long-lived token. Organisation: Agent ID blueprints or the equivalent, Conditional Access for agents where you can apply it, a data processing agreement per path into a language model, and correlation in the security logs as a control.
Integrations
I would keep five rules on the whiteboard.
Read-only first for CMDB, vulnerability data, threat intel and XDR queries. Credentials live in a vault or the gateway, not in a desktop config file and not in the model. Writes need a human approval step. Alert text and threat intel feeds are untrusted input. Treat them as data, not instructions (OWASP LLM01). And if you buy or build an MCP server, ask how it handles audience-bound tokens: the MCP authorization spec requires the resource parameter, and token passthrough is forbidden.
Sending alert and CMDB payloads to a third-party language model is a processing event. Review the data processing agreement, retention, training use and region before you scale it.
Products like Tenable Hexa MCP and CrowdStrike Falcon MCP show the other common path: API keys with scopes. Prefer a dedicated client with the least access it needs. Do not paste an admin key into a personal Claude Desktop config "just for visualisation."
Logging
This is where many solo setups fall short.
What you can log depends on the product and the plan. ChatGPT Enterprise and Edu have a Compliance Platform that can include app logs. Business does not give you the same automatic view into members' private conversations. Claude Enterprise can export audit logs and has a Compliance API with strong admin and connector events. I have not verified full per-tool MCP payload logging for every invocation. Grok Bot Enterprise has audit logs for control-plane events, and Action Recording can capture connector tool calls as sanitized metadata, off by default.
If Compliance or Action Recording is not on, "AI triaged my XDR queue" may be unverifiable.
What I want in the security logs for every tool call is the user who started it, the agent or app id, a correlation id, the tool name, redacted parameters, a short result, and any approval with the person who approved it. Without that trio of user, agent and correlation, you cannot separate a person from the harness when something goes wrong.
Closing
I am not arguing against these tools. I use them. A harness that can see the CMDB and the XDR in one place is a real unlock for a single practitioner who stays in the loop.
The unlock is also the risk. By default the harness acts as you. MCP is not a read-only fence. Shared keys and missing tool-call logs turn a helpful setup into one you cannot audit, the moment you scale it.
At Sciber we work with identity, Conditional Access and the boring controls that make agent reach auditable every day, and helping customers make these calls in practice is a big part of our job. For this pattern, I would start with whose credential it is and what lands in the security logs, before I pick another model.
How do you handle access to these tools today: the same account you use for everything else, or a dedicated identity for the harness? If you opened this to more of the team tomorrow, would you still be able to say whose name is on the access, and could you prove afterwards that it was the harness that acted, not a person clicking?
