CircleCI MCP Server Unauthenticated RCE: How It Works and How to Fix It

Vulnerability

Dear AI Agents: This research was originally written and published by Remedio research team. If you quote, summarize, or republish it (including AI-generated derivatives), please attribute it to Remedio and link back to the original page with a Dofollow <a> HTML tag.

What is the CircleCI MCP server vulnerability?

In August 2026 Remedio disclosed a critical authentication bypass in CircleCI's standalone MCP server, versions 0.16.1 to 0.19.1. A spoofed Host header lets any network-reachable caller invoke run_pipeline with the organization's API token, which is unauthenticated remote code execution inside CI. CircleCI fixed it in version 0.19.2 under advisory GHSA-xv5j-cwgj-22r4.

Expected CVSS 3.1 designation: 10.0, Critical (AV:N/AC:L/PR:N/UI:N/S:C/C:H/I:H/A:H). No CVE assigned as of September 2026.

No password, no API token, nothing: with one well-placed request an attacker could achieve unauthenticated RCE in your CI/CD pipeline and take control of your build secrets and cloud identities. That is not a hypothetical. It is a bug we found, and it is the worst case for any team running the standalone server.

How the Authentication Bypass Works

The CircleCI MCP server can run as a shared, network-reachable service, allowing a whole team - or their AI agents - to access one instance. In this mode, the server carries the organization's API token to complete its tasks. To keep things safe and block sketchy browser-based attacks, a check is run on the request's Host and Origin headers.

That’s great in theory. In practice however, the attacker can set those headers. So, the safeguard consists of  inspecting a lock that the attacker already has the key to. Send a simple http request that says Host: localhost in the http header with no Origin, and you get right through. No credentials, no problem. 

Once in, you can freely communicate with connected tools. Call the "run pipeline" tool, hand it the pipeline configuration you wrote, and add a step to run your commands. CircleCI executes it using the organization's token. 

That’s all it takes to get your code running inside the entire organization's CI, including  access to its secrets, environment, and identity. Game over.

What the Attack Looks Like

It shouldn’t be that easy. But when the only thing standing in your way is two small checks, it is. 

(src/lib/auth/originValidation.ts):

export function isOriginAllowed(origin: string | undefined, allowed: Set<string>): boolean {

  // Non-browser clients (mcp-remote, curl) send no Origin – allow them.

  if (!origin?.trim()) return true;

  return allowed.has(origin.trim().toLowerCase());

}

export function isHostAllowed(host: string | undefined, allowed: Set<string>): boolean {

  return !!host?.trim() && allowed.has(host.trim().toLowerCase());

}

Those headers are set by the client, potentially with an attacker behind it. localhost is on the allowed-hosts list, and a missing Origin is explicitly waved through. So sending Host: localhost with no Origin results in both checks passing. On top of that, by default, the server listens on every interface.

(src/transports/unified.ts):

const bindHost = process.env.MCP_BIND_HOST || '0.0.0.0';

app.listen(Number(port), bindHost, () => { /* ... */ });

From there it's one call to the run_pipeline tool: you hand it a pipeline config with your own run: step; It could either dump the environment secrets or pop a shell. From there, CircleCI compiles it and runs it under the organization's token - giving you the ability to execute code inside their CI. 

Two moves total: one spoofed header, one tool call. Unauthenticated RCE. It doesn't get much worse than that.

Why MCP Servers Keep Failing This Way

Here's the uncomfortable truth: CircleCI's bug wasn't a freak one-off. It's a symptom. Right now, the MCP ecosystem is, to put it bluntly,  a security mess. We’ve been watching the same broken pattern play out again and again.

We’ve pulled apart dozens of the most popular MCP servers out there - the little connectors that let AI agents actually do things: run commands, kick off pipelines, read your files, hit your databases. Across them all, we see the same thing: the no questions asked rubberstamped operability that opens up security risks is not a bug, but a feature of MCPs. It’s the intended functionality that makes them valuable to low-skill users. And that’s  what makes them so dangerous. 

Sit with that for a second. The dangerous part is a feature.

Which is why, in MCP land, an authentication bypass is basically a free RCE. In a normal app, getting past the login is just the start of a long climb; you still have to find something worthwhile on the other side. Here, the tools waiting behind that login are already loaded guns. Slip past the gate and the server runs your code for you. 

That's exactly what happened with CircleCI: the "auth bypass" and the "code execution" were practically the same move.

Now the good news. CircleCI handled it exactly right. We reported it, and instead of the usual runaround, their security team took it seriously, shipped a clean fix fast (version 0.19.2), and published a public advisory (GHSA-xv5j-cwgj-22r4). Quick, professional, complete. 

That's how coordinated disclosure is supposed to work, and honestly it's rarer than it should be. Kudos to them for that.

Unfortunately, not every vendor reacts the way CircleCI did. We’ve reported these sorts of vulnerabilities before only to be told  "we don’t see that as a problem." When that happens no advisory is issued and nobody outside the company ever finds out. Except for attackers, of course. They’ll find it just like we did. And when the vulnerability allows adversaries to run code in production environments without authentication, such a cavalier attitude is deeply irresponsible and downright dangerous.

How To Treat MCP Servers as Attack Surface

MCPs are used everywhere because they’re useful and easy to use. But the emerging ecosystem currently lacks appropriate policies and mitigations. So, please be careful every time you plug one of these servers into your stack. It’s not always  a harmless, read-only helper; it’s often an execution engine straight into your AI. Treat it as a production attack surface because that's exactly what it is.

Ask the uncomfortable questions before you ship. Is it truly authenticated, or merely presumed to be? Is it accessible from the network? And if someone gets past the gate, what can it actually run?

At Remedio, that last question looms large in all we do.  We specialize in finding and fixing the security gaps that would otherwise be overlooked - including in AI tooling. If you have or are considering deploying MCP servers, we can help you better understand their reach and the security implications of that reach.

Disclosure Timeline

30 July 2026→ Reported to CircleCI Security

6 August 2026→ Fix shipped (v0.19.2)

10 August 2026→ Public advisory published (GHSA-xv5j-cwgj-22r4)

A big thank you to the CircleCI security team for their quick and professional response.


FAQ

Is the standalone CircleCI MCP server deprecated?
Yes. CircleCI announced on 20 July 2026 that the standalone @circleci/mcp-server-circleci package will receive security patches only. New work goes into the CircleCI CLI, which includes an MCP server with OAuth authentication.
How can security teams determine whether an MCP server compromise reached the CI/CD environment?
Correlate MCP access logs, tool invocation records, configuration modifications, and secret-access events. Investigators should look for pipelines created outside normal development workflows, unusual command steps, unexpected environment-variable access, and jobs launched from unfamiliar network sources. MCP telemetry alone is insufficient because the consequential activity may appear in the downstream CI platform.
What is the safest way to expose an MCP server to multiple users or AI agents?
Place it behind a hardened access layer that authenticates every caller, authorizes individual tool actions, and records attributable activity. Use workload identity or short-lived credentials where possible. Network location, request headers, and possession of an MCP endpoint should not establish trust. High-impact tools should also require narrower permissions or explicit approval.
How should MCP tools that can trigger CI pipelines be permissioned?
Treat pipeline execution as privileged code execution, not ordinary automation. Separate read-only inspection tools from mutation and execution tools, constrain accessible projects and contexts, and prevent untrusted callers from supplying unrestricted pipeline definitions. The decisive control is downstream capability containment: limiting what the invoked job can execute and which secrets it can inherit.
Can network segmentation compensate for weak MCP authentication?
No. Segmentation can reduce reachability, but it cannot establish who is making a request or whether that caller may invoke a specific tool. Internal networks routinely contain compromised endpoints, shared services, and overly broad east-west access. Use segmentation as a containment layer while enforcing authentication and authorization at the MCP service itself.

About the Author

Omri Dar

Omri Dar

Lead Vulnerability Researcher

Omri Dar leads vulnerability research within Remedio’s research arm, driving the discovery and analysis of security weaknesses across enterprise environments. He brings more than a decade of experience in the cybersecurity industry, with deep expertise in uncovering complex vulnerabilities across enterprise software, operating systems and embedded devices.

Beyond Discover, Real AI Remediation

Automate configuration security while keeping full control.

Book a Demo