governance
Keep ChatGPT and Claude, add a guardrail: least-privilege AI execution
A walkthrough for IT and security reviewers: how to keep ChatGPT and Claude while bounding what they can reach, from the connector model and ChatGPT setup to roles, permissions and one governed run.
- #governance
- #mcp
- #ai-agents
- #tool-use
- #roles
Banning ChatGPT and Claude is the reflex of many security teams, and the argument behind it is real: an assistant wired to company systems can read and act far beyond what any single employee was granted. What the ban does not explain is why it rarely holds. In a January 2026 survey of 2,000 employees in the UK and US, BlackFog found that 63% think it is acceptable to use AI tools without IT oversight when no approved option exists, and 51% have already connected AI tools to other work systems without IT approval.
The assistant is not the risk. The ungoverned connection is. This walkthrough is for the IT or security reviewer who has to approve the connection: how the connector model works, how ChatGPT is connected step by step, how permissions map to skills and roles, and what one governed employee run looks like.
What is least-privilege AI execution?
Least-privilege AI execution means an AI assistant can reach only the tools and data one task needs, with the rights of the person who asked and never more, under rules enforced outside the model. The employee keeps their usual assistant. What changes is what the assistant is allowed to call.
The principle is old. Saltzer and Schroeder wrote it down in 1975, and it is the core of the role-as-permission-boundary model for agents. The new part is where it has to be enforced. OWASP lists the failure as LLM06:2025 Excessive Agency: an agent with more functions, more permissions or more autonomy than its task needs.
Why keep the assistants employees already use?
Because the alternative is a second, unofficial set of assistants. People reach for the tool that works, and a ban moves usage to personal accounts, where no company control applies. That pattern is shadow AI, and its sibling for connectors, shadow MCP, is what happens when employees plug servers into their assistant themselves.
Ali Mahfouz, Area Service Lead Romandie and member of the executive committee of Microsoft Switzerland, put the stance plainly at Vision Entreprise 2026 (University of Lausanne and PME magazine, 8 October): “Guardrails are not there to slow adoption down. On the contrary, they are there to make deployment at scale possible.”
How does the connector model enforce least privilege?
The assistant connects once, over the Model Context Protocol (MCP), to skilder rather than to each business system. skilder then serves it roles: bundles of skills and scoped tools drawn from an internal, curated registry. The assistant only ever sees the tools of the role in use, loaded when a skill needs them.
Each layer bounds something different, and each has a different owner.
| Layer | Who decides | What it bounds |
|---|---|---|
| Assistant connection (OAuth for MCP, enterprise SSO) | IT administrator | Which assistant may talk to the workspace, and as whom |
| Internal MCP registry | IT and security | Which business systems exist as tools at all |
| Role, granted through role-based access control | Role owner and administrator | Which tools a team’s assistant can see |
| Per-user authorisation on a connector | Each employee | Data the person can already reach, never more |
| Guardrails (beta) | Security or the role owner | Which individual calls pass, before or after they run |
Three properties do most of the work for a reviewer.
Tools appear on demand, inside the role. An assistant asking for help first gets the matching role, then that role’s skills, and only then the tools those skills use. A tool outside the role never enters the conversation, so it cannot be called by a planner that took instructions from a supplier PDF.
The requester’s rights, not a shared key. Where a connector needs the user’s own authorisation, the call is refused until that person has connected their own account in skilder. The assistant acts with the employee’s access to the CRM or the file share, never with an administrator token that happened to be at hand.
Read and write are separate. Tools are classified as read or write, and guardrails, in beta, can cover them by server, by tool name, by kind and by role. A guardrail is a rule written as a script (Python, JavaScript, TypeScript or Bash) that skilder checks on each call, in observing or blocking mode. The assistant cannot talk its way past it, because the model never evaluates it.
How do you connect ChatGPT step by step?
On ChatGPT Business, Enterprise and Edu, custom MCP connections are called apps and run in developer mode. On Business plans, only admins and owners can enable developer mode and publish an app, according to OpenAI’s help article in October 2026. OpenAI moves these menus regularly; the labels below follow that article.
- Enable developer mode. In Workspace settings, open Permissions & Roles, then Connected Data, and turn on developer mode for creating custom MCP connectors. Each admin enables it for their own account.
- Create the app. In Workspace settings, open Apps, then Create.
- Point it at skilder. Enter the skilder MCP endpoint of your workspace and choose OAuth as the authentication method.
- Sign in. ChatGPT opens the skilder sign-in. Use the company’s single sign-on; the connection is tied to that identity.
- Test it privately. In a new chat, ask for a task the target role covers and check which role and tools come back.
- Publish it for the workspace. Once published, the app appears in the workspace’s approved list. On Enterprise and Edu, role-based access control in ChatGPT decides which members can use it.
- Read the trace. Open the call log in skilder and confirm the test run is attributed to your account and role.
Claude and Copilot Studio connect to the same endpoint over MCP; their admin screens differ, the model behind the connection does not. OpenAI’s own note applies to every client: connecting an untrusted MCP server increases exposure to prompt injection, which is why the registry behind the endpoint matters more than the assistant in front of it.
How do permissions map to skills and roles?
A skill is a team’s know-how made executable: the method, the tools it needs and the format of the result. A role groups a trade’s skills, their scoped tools and their rights. Roles are configured from a blueprint against the team’s own processes, then granted to people.
| Question a reviewer asks | Where the answer lives |
|---|---|
| Who may use this capability? | The role, granted through role-based access control |
| Which systems can it touch? | The tools attached to the role’s skills, from the internal registry |
| With whose rights? | The requester’s own authorisation on each connector |
| Who changes the method? | The role owner, who edits drafts and publishes versions |
| What did it do last Tuesday? | The execution trace, per person, per role, per call |
Publishing is a deliberate step. A skill is edited as a draft, tested in Studio in a real conversation that shows which skills and tools the role used, then published. Automated evaluations that re-run tests on publish are coming soon; until then, testing is manual.
What does a governed employee run look like?
A sample run, with sample data: Talvero Bau AG is a fictitious construction company.
- A site manager asks ChatGPT for the weekly report on the Sion depot extension.
- ChatGPT asks skilder which role fits and receives Site Reporting, the only role this manager holds that covers the request.
- The report skill loads its method and two read tools: the site diary and the photo archive. No mail tool is in reach.
- The assistant lists the week’s diary entries with the manager’s own rights. Entries from another site, which the manager cannot see, never come back.
- The manager asks ChatGPT to send the report straight to the client. The role has no send tool, so the assistant drafts the email for the manager to send.
- Every call lands in the trace: the person, the role, the skill version, each tool with its input and output.
Nothing in that run depends on the prompt holding. The boundary sits in the role, the connector and the guardrails, outside the model.
What should the security review check?
A short list to run before approving the connection:
- One endpoint, one registry. The assistant connects to a single governed MCP endpoint, and every tool behind it comes from a reviewed list.
- No shared tokens. Connectors use the requester’s own authorisation, or a service credential scoped to the role, never a personal admin token.
- Write actions are rare and named. Each write tool belongs to a role that needs it, and the role says so in one sentence.
- Guardrails on the risky edges. Outbound email, external sharing and deletes get a rule, in observing mode first, then blocking.
- Owners for every role. A named business owner publishes changes, and versions are kept.
- The trace answers “who did what”. A test run shows person, role, skill version and each call before production data is connected.
To go through this checklist against your own assistants and systems, a free 30-minute exchange with an expert is the shortest route. For how roles compare with agent plugins as a packaging unit, see agent plugins vs. skilder roles.
Takeaway: keep ChatGPT and Claude, and move the control to where the model cannot argue with it: the role, the connector and the rule on each call.
To review how your assistants should connect to your business systems, book a free 30-minute exchange with an expert.
Frequently asked questions
Can we allow ChatGPT at work without losing control of company data?
Yes, if control sits in the connection rather than in a ban. Use a business plan with an admin-managed workspace, connect business systems through one governed MCP endpoint instead of personal connectors, scope tools per role, and keep an execution trace. A January 2026 BlackFog survey found 63% of employees accept unsanctioned AI when no approved option exists, so an approved path matters.
How do I connect internal business systems to Claude and ChatGPT securely for employees?
Connect each assistant once to a governed layer over the Model Context Protocol, rather than wiring every system into every assistant. The layer exposes only reviewed tools from an internal registry, grants them per role, uses each employee's own authorisation on connectors, applies rules to individual calls and records who ran what. Claude, ChatGPT and Copilot Studio can share the same endpoint.
Who can add a custom MCP connector in ChatGPT Business?
On ChatGPT Business, only workspace admins and owners can enable developer mode and publish a custom MCP app, according to OpenAI's help documentation in October 2026; members cannot be given developer mode individually. Enterprise and Edu workspaces add role-based access control to choose who may build apps and who may use each published app.
What is least privilege for AI agents?
Least privilege for AI agents means an agent can reach only the tools and data one task requires, with the rights of the person who asked and nothing broader. OWASP's Top 10 for LLM applications names the opposite failure LLM06:2025 Excessive Agency: an agent with more functions, permissions or autonomy than the task needs. The role is the practical unit for setting that boundary.
Does connecting skilder give ChatGPT access to all our tools?
No. ChatGPT sees the tools of the role in use, loaded when one of that role's skills needs them, and those tools come from the workspace's internal, curated registry. Where a connector requires the user's own authorisation, calls are refused until that employee has connected their account, so the assistant acts with that person's existing access only.
What happens when an AI agent tries an action its role does not allow?
If the tool is not part of the role, the agent never sees it and cannot call it. If the tool is in reach but a guardrail rule covers the call, skilder checks it before or after it runs and, in blocking mode, refuses it. Guardrails are in beta; refused calls are logged with the rule's name and its reason.
Is a guardrail the same as an instruction in the prompt?
No. A prompt instruction asks the model to behave, and a prompt injection in a document can override it. A skilder guardrail is a rule written as a script that the platform runs on the tool call itself, outside the model, so the assistant cannot negotiate with it. Guardrails can run in observing mode first, then in blocking mode.
Related articles
-
October 1, 2026
Progressive skill discovery as access control: what the skilder paper measures
A 30-page white paper tests whether serving skills through one MCP server can enforce what an agent may do. 13 scenarios, six models, three tool interfaces, and the limits the authors report.
-
October 23, 2026
The future of work with AI: what Anthropic's experience can teach companies starting to adopt AI
Anthropic now runs on AI agents. Most companies do not, and do not need to. What its 2026 data teaches about the future of work with AI, and how to get ready one workflow at a time.
-
April 20, 2026
The Context Lake: Why Your Data Lake Isn't Enough for AI Agents
A new layer is taking shape in the agentic enterprise stack: the context lake. Business context, tool permissions, and governance need their own home.