Where the risk actually sits
MCP itself is a way for an AI tool to call a server. The risk comes from three places: what the server can reach, whose rights it uses, and what text flows back into the model. The MCP project’s security best practices (for the 2026-07-28 specification) list the attacks that matter:
| Risk | What happens | Control |
|---|---|---|
| Token passthrough | A server accepts or forwards tokens not issued for it | Only accept tokens issued for this server; the spec forbids passthrough |
| Confused deputy | An attacker abuses an MCP proxy with a fixed client ID to reach a backend service without valid consent for that client | Per-client consent; Dynamic Client Registration is deprecated in this spec |
| SSRF | A server is tricked into calling internal addresses like 169.254.169.254 | Block private and link-local IP ranges, require HTTPS, use an egress proxy |
| Session or state hijacking | An attacker reuses a handle or session | Bind state to the user, validate every request |
| Local server compromise | A start command runs malicious code | Show the exact command first, sandbox, prefer stdio with limited rights |
| Prompt injection | Tool output contains instructions | Limit accessible data and destinations, treat tool content as untrusted, and approve consequential actions. Read-only access does not prevent data disclosure |
Controls a company should set
- Trusted servers only. Claude says custom connectors are not verified by Anthropic; OpenAI says it does not review custom servers. Someone in your company decides which servers are allowed.
- Admin control per tool. On Claude Team an Owner or Primary Owner adds connectors; Enterprise also allows custom roles with Manage access to Libraries. On Copilot Business and Enterprise the MCP policy is off until an admin enables it. Use these switches.
- Per-person sign-in, not shared keys. Claude warns that a shared API key on Team or Enterprise gives every user the same access. Prefer OAuth so each person acts with their own rights.
- Minimal scopes. The MCP guidance says start with minimal, read-level scopes, raise them step by step, avoid wildcard scopes and log every elevation.
- Read-only first, approve writes. Check ChatGPT’s app permissions and confirmation rules. GitHub advises read-only tools. Claude advises “Allow always” only for servers and tools you trust without oversight, and suggests disabling write tools during Research, where tools can run without approval.
- Limit tools per server. Use
toolsin Copilot andallowed_toolsin the xAI API. xAI does not supportrequire_approval, so this is your main lever there. - Watch delegated agents. Copilot cloud agent and code review use configured MCP tools without asking for approval when you delegate work to them.
- Monitor. Server behavior can change after you approved it. Log tool calls and review them.
The misunderstanding
“It runs locally, so our data stays safe.” A local server runs with your rights and can run code on your machine. And whatever it returns still goes to the model you use. Local is a choice about where the server runs, not a security model.
What this means for a team
Every AI tool has its own MCP settings, admins and approval rules, so a policy that only covers Claude leaves ChatGPT and Copilot open. Write one list of allowed servers and scopes, and apply it in each tool. For data rules beyond MCP, see is ChatGPT safe for business data (Dutch) and how to prevent AI mistakes and review its work. Setup per tool is in how to connect MCP to Claude, ChatGPT, Copilot and Grok.
Where Bombos fits
Bombos is a memory plugin, so it is one more thing your AI tools connect to, and it should meet the same bar. It gives every AI tool a team uses one shared memory with team knowledge, learned rules and what already sits in mail, documents and repos, shows what each answer rests on, and can run on your own server with your own models. One person can try it first with their own agents before IT decides on the team.
Related questions
What is prompt injection in MCP?
Text inside a tool result, such as an email or a web page the server fetched, that tries to give the AI new instructions. Claude, OpenAI and GitHub all name it as a core risk of connecting MCP servers.
Are local MCP servers safer than remote ones?
Not automatically. A local server runs with the rights of the client on your machine, and VS Code warns that local servers can run code on it. The MCP security guidance advises showing the exact start command before running it and using a sandbox.
Can we turn MCP off for our organization?
In GitHub Copilot Business and Enterprise, the MCP servers in Copilot policy is off by default and admins decide. In Claude Team and Enterprise only owners (and on Enterprise, roles with Manage on Libraries) can add custom connectors.
Are custom MCP servers automatically reviewed by Anthropic or OpenAI?
No. A custom connection is not automatically vendor-reviewed; this differs from reviewed integrations in official directories. Claude's help center says custom connectors are not verified by Anthropic, and OpenAI says it does not build or review custom MCP servers.
