Guide · AI security · as of 13 November 2026

An MCP server could send one client’s answers to another. Four checks for yours.

MCP, the Model Context Protocol, is how an AI assistant such as Claude or ChatGPT connects to your own systems: a CRM, a database, a document store. The small program in the middle is called an MCP server. On 5 November 2026 the US National Vulnerability Database listed two flaws in the official TypeScript kit that many of those servers are built with. Neither needs a clever attacker. One could hand a client the answer meant for someone else during normal use. The other accepted a login issued for a different service. This guide works from the official records and turns them into four checks.

Flaw one

CVE-2026-111770

One shared server, answers crossed.

The record says that in the server package from version 2.0.0 until 2.3.0, an application that shared one server object among several clients over HTTP could deliver responses and notifications to the wrong client. The cause is simple. The kit matched each answer to a connection using only the request number, and its clients start counting at zero. Two clients asking at the same moment can both be on request number 3. The record says these collisions occur in normal operation. It is scored 7.7 out of 10.

Flaw two

CVE-2026-111771

A valid pass, for another door.

The second record covers the kit’s login check for HTTP servers. It did not confirm that an access token was issued for this server. Where one login system issues tokens for several services, a token for another service was accepted here too. It is scored 6.8 out of 10. The fix has a catch that is easy to miss. The record says upgrading alone leaves the new check switched off. You must name the audience you expect and also return the token’s audience from your own verify step. Set only the first and every request is refused.

Four checks for any MCP server you run

Ask whoever built your connector.

  • 1. Read the version. The fixed versions are server 2.3.0, sdk 1.32.0, express 2.0.2 and server-legacy 2.3.1. Test: open the lock file of the project and read the installed version of each package.
  • 2. Build a new server for each request or each session, never one at start-up. After the upgrade a shared server fails on its second request, so plan the code change with it. Test: send two requests from two clients at once and confirm each gets only its own answer.
  • 3. Check who the token was issued for. If you use the kit’s bearer login, switch on the audience check and return the audience from your verify step. Test: present a valid token from another service and confirm a 401.
  • 4. Limit what a wrong answer can cost. Give the connector read-only access where you can, and only the tables it needs. Test: list every tool the server offers and what data each can return.

What we checked on our own

We run one. So we looked.

We build MCP connectors, so these records were about our own work. On 13 November 2026 we checked the connector we run in production. It is on server version 2.4.0, past the fix. Its handler builds a new server for every request, the pattern the advisory lists as not affected. It does not use the kit’s bearer login, so the second flaw does not apply to it. The check took minutes because the connector is small and read-only. That is the lesson we would pass on: an AI connector is a server like any other, and somebody has to own its updates.

AI agents for Dubai businesses

Questions

Questions about MCP server security.

  • What is an MCP server?

    It is a small program that lets an AI assistant read from or act on one of your systems through the Model Context Protocol. When you add a custom connector to Claude or ChatGPT, you are pointing it at an MCP server.

  • Is my MCP server affected?

    For the first flaw, only if it is built on the TypeScript server package from 2.0.0 until 2.3.0, is reached over HTTP by more than one client, and builds its server once at start-up. Servers that run over stdio on one machine are not affected. For the second, only if it uses the kit’s bearer login and your login system issues tokens for several services.

  • Is upgrading enough?

    No, in both cases. For the first flaw the upgrade makes a shared server fail, so the code must change to build one per request or session. For the second the audience check stays off until you switch it on and supply the audience from your verify step.

  • Has anyone been attacked through these flaws?

    The records we read do not report attacks, and neither flaw was on the CISA list of exploited vulnerabilities when we checked on 13 November 2026. The first flaw does not need an attacker to cause a leak, which is the reason to fix it anyway.

Sources

read 13 November 2026

Check it at the source.

  1. NVD: CVE-2026-111770, responses delivered to the wrong client (the description, the score and the fixed version)
  2. NVD: CVE-2026-111771, tokens for another service accepted (the affected packages and the two-part fix)
  3. MCP TypeScript SDK: advisory GHSA-jm63-cq47-j5jp (who is affected and who is not)
  4. GitHub Advisory Database: GHSA-rvq5-wwqv-78pq (the audience setting and how to turn it on)

A question?

Connecting AI to your own systems?

Tell us which assistants your team uses and what they should reach. We will build the connector, keep it read-only where it can be, and own its updates.

Talk to usWhatsApp +971 50 728 5494hello@brevdigital.com