AWS has disclosed a high-severity authorization flaw in its open-source MCP registry: any authenticated user could grant themselves administrator rights over every registered MCP server, and read the credentials the registry stores for them.
What happened
On 11 November AWS published security bulletin 2026-179-AWS for CVE-2026-100307, a missing-authorization issue rated 8.8 (high) on the CVSS 3.1 scale. The affected product is Amazon mcp-gateway-registry, the open-source registry and gateway for Model Context Protocol servers that centralises discovery, access control and proxying of registered MCP tool assets.
The /api/servers/* management endpoints checked that a caller was authenticated but never checked role or permission. A low-privileged user could send a crafted group import to POST /api/servers/groups/import that mapped their own identity-provider group to the mcp-registry-admin scope — or named a privileged scope directly, or granted a mutating UI permission on all servers. The registry wrote that group definition to its scope store and derived administrator status per request, so the escalation took effect on the caller’s next request, with no re-authentication.
The result was administrative control of every registered gateway asset. An attacker who reached that state could register, modify, toggle and remove servers and agents; read and change access-control groups; and read the backend credentials and configuration the registry manages, including stored authentication credentials and custom headers, which the registry decrypts on read through administrator-reachable routes.
Affected versions are 1.0.9 through 1.24.7. Version 1.25.0, released in late June, enforces authorization on the write path. AWS published the advisory on 11 November; the GitHub advisory GHSA-rxf3-qgpv-wxvr, published the same day, carries the same affected range and credits the researcher novice-22.
Upgrading is not the whole fix, and AWS says so explicitly. The patch does not rewrite group definitions written before the upgrade, and it does not invalidate credentials read before it. A privileged mapping created through the vulnerable endpoint still confers administrator access afterwards. AWS’s post-upgrade checklist asks operators to inspect every group_mappings entry in the mcp_scopes collection for a mapping to mcp-registry-admin or another privileged scope, and every ui_permissions entry granting a mutating action to all, removing any grant an administrator did not create; to search HTTP logs for POST /api/servers/groups/import from non-administrator accounts, along with the sibling group routes and server-mutation routes, during the period an affected version was deployed; to rotate registry-managed credentials where unintended access cannot be excluded; and to review registered assets for changes in that window, including upstream URLs pointed at a different host.
For deployments that cannot upgrade yet, AWS recommends restricting the management API at the ingress or proxy layer to trusted administrator identities, and monitoring the scope store for new privileged mappings.
Why it matters for KYA
A registry is the part of the agent stack that says which tools exist and who may reach them. That makes its own authorization model part of the trust boundary. This is a clean demonstration of what happens when the model is missing: the authentication check held, the authorization check did not, and the gap handed out control of the directory itself.
The blast radius is larger than a typical privilege-escalation bug because of what a registry knows. An MCP gateway holds the credentials agents use to reach registered servers. An attacker who becomes registry administrator can read those credentials, repoint an upstream URL at a host of their choosing, and register new servers into the catalogue that agents will then discover and call. Verifying the agent in front of the tool is worth little if the catalogue behind it can be rewritten.
The persistence problem is the part KYA practitioners should take away. Patching closed the write path; it did not undo grants already written, and it did not revoke credentials already read. That is the same asymmetry that runs through agent identity generally: revocation has to reach the artifacts that were issued, not just the code path that issues them. An audit that only checks current configuration will report a clean registry while a stale administrator mapping sits in the scope store.
What to watch
- Whether other agent registries and MCP gateways publish comparable authorization reviews. The registry pattern is spreading faster than its threat models.
- Whether registry products move to per-request authorization with auditable grant records, so a mapping change becomes a logged event rather than a silent write.
- The gap between fix and advisory here — roughly four and a half months — and what it means for operators who treat “we are on the latest version” as evidence of safety.
Sources
- AWS Security Bulletin 2026-179-AWS, “CVE-2026-100307 — Missing authorization on the external management API in mcp-gateway-registry enables privilege escalation to registry administrator”, 11 November 2026 — aws.amazon.com
- GitHub Security Advisory GHSA-rxf3-qgpv-wxvr, “Missing authorization on /api/servers/* enables privilege escalation to gateway admin”, 11 November 2026 — github.com
- CVE Program record CVE-2026-100307, CVSS 3.1 base score 8.8 — cve.org
- mcp-gateway-registry release 1.25.0 — github.com
Know Your Agent (KYA) explains agent identity, verification and accountability. This is an explainer, not legal or compliance advice — see our Sources & methodology. New to KYA? Start with What is Know Your Agent? and the glossary.