Cybersecurity

AWS MCP Gateway Registry SSRF Flaw: Update to 1.35.0, Then Audit Your Servers

Published By Shivam Sharma7 min read
AWS MCP Gateway Registry SSRF flaw CVE-2026-125926 — update to version 1.35.0

Amazon Web Services has disclosed a server-side request forgery flaw in mcp-gateway-registry, its open-source registry and gateway for Model Context Protocol servers, that let an authenticated user point the gateway at services on the deployment’s internal network, including the cloud instance metadata service that issues credentials. The flaw, CVE-2026-125926, affects versions 1.0.3 through 1.34.0 and is fixed in version 1.35.0. The AWS MCP Gateway Registry SSRF is another similar flaw AWS has disclosed in one of its own AI-agent platforms in about six weeks, and the fix comes with a catch: upgrading alone does not clean up configurations an attacker may already have stored.

How the flaw works

mcp-gateway-registry centralizes discovery, access control and proxying of registered MCP tool assets. When a server is registered or updated, the API accepts an upstream address that the gateway may later use when sending requests. The flaw is that the registry did not validate the upstream address at all. An authenticated user with permission to register or modify a server could set a crafted upstream address, and the gateway would then send requests to services on the deployment’s internal network, services the user could not reach directly. In some cases the gateway returned those responses to the user.

The credential path is what makes this serious. Depending on network configuration, the reachable targets can include the cloud instance metadata service and the credentials it issues. On Amazon EC2, the metadata service is at the link-local address 169.254.169.254 and can provide credentials. Depending on network configuration, an attacker who can make the gateway reach the metadata service may be able to access credentials it issues.

Two limits matter. First, this is not an unauthenticated flaw: the attacker needs a valid account with permission to register or modify servers. Second, there is no confirmed exploitation. As of 14 November, no public proof of concept and no in-the-wild exploitation had been reported, the flaw is not on CISA’s Known Exploited Vulnerabilities list, and exploitation-probability scoring puts it at about 0.3 percent, according to Threadlinqs. None of that makes the flaw safe. It makes it a known hole with a known fix, in a product whose whole job is brokering access to AI tools.

A second flaw in the same product, disclosed the same day

AWS published two bulletins for mcp-gateway-registry on Wednesday 11 November. The SSRF is bulletin 2026-180-AWS, published at 13:00 PST. The second flaw, CVE-2026-100307, is bulletin 2026-179-AWS, published at 12:30 PST the same day. It is a missing-authorization issue in the external management API: the /api/servers/* endpoints authenticated the caller but performed no role- or permission-based authorization, so an authenticated low-privileged user could send a POST to /api/servers/groups/import and map their own identity-provider group to the registry administrator scope. The escalation took effect on the caller’s next request, with no re-authentication, granting administrative control over every registered gateway asset.

The second flaw affects versions 1.0.9 through 1.24.7 and was fixed in version 1.25.0, released 27 June 2026. AWS credits external researcher novice-22 with the coordinated disclosure of the SSRF flaw. Theinjection.dev says novice-22 is credited for both flaws. The GitHub advisory for the SSRF is GHSA-fc39-fwh3-wv5r.

The 13-day gap and the 124-day window

The timeline shows a gap between the fix and the public bulletin. Version 1.35.0, the release that fixes the SSRF, was published on Thursday 29 October 2026. AWS’s public bulletin did not go out until Wednesday 11 November, 13 days later. For 13 days, the patched version existed before AWS published its public security bulletin.

The second flaw creates a longer trap. Version 1.25.0, which fixed the admin escalation, shipped on Saturday 27 June 2026. The SSRF fix, 1.35.0, shipped 124 days later, on 29 October. A deployment running 1.25.0 or later had the admin-escalation fix but remained affected by the SSRF until it was upgraded to 1.35.0 or later. The two releases were 124 days apart. AWS flags this sequencing trap in its own bulletin: a deployment running 1.25.0 or later carries the admin-escalation fix but remains affected by CVE-2026-125926 until it runs 1.35.0 or later.

Why upgrading alone does not fix the AWS MCP Gateway Registry SSRF

Version 1.35.0 validates upstream addresses when a server is registered or modified. It leaves server configurations stored before the upgrade unchanged. That means an unintended or malicious upstream address stored while an affected version ran remains in the configuration after the patch. The same applies to the admin-escalation flaw: upgrading to 1.25.0 or later stops new abuse but does not remove privileged mappings created while an affected version ran. Those must be checked in the mcp_scopes collection.

AWS tells customers who ran an affected version to take three steps after upgrading. First, review the upstream addresses of registered servers, and correct or remove any address you did not intend. Second, review application logs from the period an affected version was deployed for unexpected server registrations and changes. Third, if the registry runs with an attached cloud role and the log review cannot rule out unexpected activity, review that role’s activity and rotate its credentials.

For deployments that cannot upgrade yet, AWS lists three workarounds, and says plainly that they do not replace the post-upgrade actions: limit the permission to register and modify servers to trusted accounts; restrict outbound network access from the registry and auth-server components to the destinations they need; and on Amazon EC2, require IMDSv2 and grant each component’s role only the permissions it needs. AWS also says any forked or derivative code must be patched to incorporate the new fixes.

What should I do if I run mcp-gateway-registry?

Upgrade to mcp-gateway-registry 1.35.0 or later, then audit every stored server’s upstream address and your application logs. Upgrading alone does not clean up old configurations. If your registry runs with a cloud role attached, review that role’s activity and rotate its credentials if the logs cannot rule out unexpected access.

  1. Upgrade every deployment to mcp-gateway-registry 1.35.0 or later, including any forked or derivative code.
  2. Review the upstream address of every registered server, and correct or remove any address you did not intend.
  3. Review application logs from the whole affected period for unexpected server registrations and changes.
  4. If the deployment also ran an affected version of CVE-2026-100307, check the mcp_scopes collection for privileged mappings from that period.
  5. Review attached cloud-role activity and rotate its credentials if the logs cannot clear it.

Steps 1, 2, 3 and 5 follow AWS’s post-upgrade guidance in bulletin 2026-180-AWS. Step 4 is an additional check from bulletin 2026-179-AWS, for deployments that also ran an affected version of the second flaw. If you cannot upgrade yet, apply the workarounds above, and treat them as a stopgap, not a fix.

The CVSS number, explained

If you saw two different severity scores for the AWS MCP Gateway Registry SSRF, both are real, and the difference is the scoring version. The GitHub advisory scores CVE-2026-125926 at 8.5, High, under CVSS 3.1, with the vector AV:N/AC:L/PR:L/UI:N/S:C/C:H/I:L/A:N. NVD’s CVSS 4.0 score is 6.3, Medium. The 2.2-point gap is not a disagreement about the flaw. It is the same flaw measured by two versions of the scoring system, and the older 3.1 number is the one the advisory itself carries. AWS rates the bulletin “Important”. For the second flaw, CVE-2026-100307, a CVSS 8.7 score comes from Corgea’s roundup only, so treat that number as Corgea’s, not AWS’s.

The pattern AWS keeps repeating

This is another time in about six weeks that AWS has fixed similar SSRF and authentication flaws in one of its own AI-agent platforms. On Friday 2 October 2026, bulletin 2026-124-AWS disclosed similar SSRF and authentication flaws in Loom, AWS’s agent platform, where authenticated users with write scopes could redirect backend connections to internal endpoints, including container-credential endpoints. The Loom flaw is CVE-2026-103958, also CWE-918. Forty days separate the Loom bulletin from the mcp-gateway-registry bulletin.

Both cases involve authenticated users redirecting platform connections toward internal endpoints. The implementations differ, but the shape is the same: an input the platform trusts, pointing it somewhere it should not go. A separate academic study of 273 MCP servers found 7 SSRF cases, alongside 64 path traversal, 5 data exfiltration and 3 OS command injection findings, according to the nudefend blog. That study concerns MCP servers in general, not mcp-gateway-registry deployments. It recorded seven SSRF cases among the 273 servers examined. For this flaw, the key fix is to validate upstream addresses whenever a server is registered or modified.

What to watch next is straightforward. Whether any exploitation surfaces now that the flaw is public, whether AWS revises either bulletin, and whether other MCP gateway and registry products get the same treatment. Until then, the message for anyone running mcp-gateway-registry is the one AWS wrote itself: upgrade to 1.35.0, then audit what is already stored. The audit matters because the upgrade does not change server configurations stored before the patch.