Loom for AWS, per-user MCP keys follow an endpoint move
An mcp:write admin inside a group they already administer can move an MCP endpoint_url. Loom deletes the admin API key and leaves per-user keys at loom/mcp/{name}/api-key/{sub}. A later invoke sends the personal key to the new host. Same group only. Labbed on v1.9.0. The helper is unchanged on v1.10.3, which I re-read and did not rebuild.
- Name
- Loom for AWS, per-user MCP keys follow an endpoint move
- Type
- 0-day analysis
- CVE
- n/a
- CVE Risk
- medium
- Disclosure Status
- public
- Vendor
- Loom
- Affected
- Loom for AWS v1.9.0 (a3050669f5b4c83210a44ed4346f17e01641e82f), labbed in one group, and v1.10.3 (cc10ab08573087fa12154f7c8181ffcd985fea88), source re-read. Caller holds mcp:write on g-admins-demo. Not cross-group.
- Published
- 16 Nov 2026
- Updated
- 16 Nov 2026
- Tags
- 0-day, loom, aws, cwe-522, mcp
The admin key left. The personal key did not.
They clear the shared key when an MCP server changes address. They do not clear the personal ones, and the next invoke sends those to the new host.
The pack is abraxas/loom-mcp-userkey-follow. The loopback client is loom-mcp-userkey-follow-Abraxas-Labs.py. I post as @abraxas_null.
Inside one team, the person who manages connectors can change where a connector points. Loom throws away the shared key when the address changes. It keeps each person's own key. The next time someone asks an agent to use the connector, that personal key is sent to the new address. My run used one account for the change and for the chat. A second person's key is the same lookup, and it leaves only when that person uses the connector. This stays inside the team. It is not a hop to the instance metadata service, and it is not code execution.
Loom is the control plane I was reading after the published bugs were already closed and the MCP secret paths had been rewritten more than once. CVE-2026-103956 (unauthenticated super-admin, fixed in v1.6.1), CVE-2026-103957 (OAuth discovery), and CVE-2026-103958 (MCP and A2A SSRF) are bulletin 2026-124. I am not refiling them. v1.10.1 closed the sibling I wanted this to be, plus three others: planting a secret ARN in an agent config, deleting a credential provider through that config, and harness invoke binding an MCP server by display name. Those four are shut. This page is the key the endpoint move forgot.
Medium 6.9, CVSS:3.1/AV:N/AC:L/PR:H/UI:R/S:C/C:H/I:L/A:N, CWE-522. Remote only in the sense that it is the HTTP API. The changed scope is the personal key leaving for a host the mover picked. Integrity is low because mcp:write may already edit the URL. The key on the wire is the part outside that grant. UI:R is someone else invoking later. My container did not need that click. Availability was not touched.
I ran the container on tag v1.9.0, commit a3050669f5b4c83210a44ed4346f17e01641e82f. The re-read was tag v1.10.3, commit cc10ab08573087fa12154f7c8181ffcd985fea88. Same helper, same clear that skips per-user names. I did not boot v1.10.3 for this finding.
The caller is g-admins-demo, which includes mcp:write and invoke. The row was tagged loom:group of demo, the group this caller already administers. Same group only. The closed bug next door is a demo admin stamping a row into someone else's group. This is not that, and I am not going to borrow its impact.
The name is the secret
user_api_key_secret_name returns loom/mcp/{server_name}/api-key/{user_sub}. The comment says the path is embedded in AGENT_CONFIG_JSON at deploy time and read by the deployed agent, so re-keying it on server id would strip personal keys until somebody redeployed. It then says the name is safe to leave in the path because cross-group collisions are refused, and because the trailing subject confines each entry to its owner.
The first half of that argument is real. I am not claiming a cross-group rename. The second half is the hole. The subject says whose key it is. It does not say which host may receive it. The person who can change endpoint_url never has to guess the secret's name. Invoke looks the name up.
A personal key, if the phrase is muddy, is the token someone stored so an agent can call a tool as them. Moving the tool's address should make that token useless until they type a new one. The admin key got that treatment. The personal key did not.
The move is in the MCP update handler, here. moved_endpoint means the URL string changed. If the OAuth flag is the string true, the id-keyed OAuth client secret is deleted. If the admin flag is true and the request did not supply a replacement api_key, the id-keyed admin secret and the legacy name-keyed admin secret are deleted, and has_admin_api_key becomes false. The comment on that block is the right threat: a credential for one host must not follow the record, because mcp:write is enough to repoint a server. Then the function returns. It never calls user_api_key_secret_name, and it never lists loom/mcp/{name}/api-key/*.
POST /api/agents/{id}/invoke with connector_ids loads those rows and, for api_key auth, sets the secret name from user_api_key_secret_name(server.name, user.sub) and the live endpoint_url. The header is the row's api_key_header_name, or x-api-key if that column is empty. I set it to Authorization, which is why the oracle recorded a Bearer token. The value is the personal key either way. The admin key was not on that request.
Where I was wrong
The finding I sat down to prove was the cross-group plant. v1.10.1 closed it. A rename onto a name held in another group is a 409. This run never left demo.
I assumed the move wiped every secret. It wipes the admin key, and the OAuth client secret when the flag equals true. Per-user keys are not columns, so the update path never lists them. After the PUT the admin secret was gone and the flag was false. The user secret was still the witness. The clear works. It works on the wrong set.
guarded_post still blocks loopback, the metadata addresses, and 0.0.0.0/8. It allows RFC1918 on purpose. The URL I stored was a private lab address. On the host the capture process was 127.0.0.1:18621 and the product was 127.0.0.1:18620. The product was not told to fetch 127.0.0.1. Not a new SSRF, and not CVE-2026-103958.
I invoked as the same principal who stored the key. One subject, one header, the witness in it, the admin value absent. That is not a second captured header. Another user's key is the same pattern with their subject, resolved when they invoke. I did not have that person in the container. UI:R is their later click. The lab line is the same-admin case.
What an attacker can do
| Method | Class | What they get | Proof |
|---|---|---|---|
POST /api/mcp/servers in demo, auth type api_key |
integrity | A connector this caller was already allowed to create. | Lab-proven, 201, admin flag set. Not an escalation by itself. |
PUT /api/mcp/servers/{id}/api-key |
confidentiality, staged | Stores the caller's own key at loom/mcp/{name}/api-key/{sub}. |
Lab-proven. The value was the witness. |
PUT /api/mcp/servers/{id} with a new endpoint_url and no new api_key |
integrity | URL changes. Admin flag false. Admin secret deleted. User secret remains. | Lab-proven. Move 200. |
POST /api/agents/{id}/invoke with that connector id |
confidentiality | The new host receives the personal key. Here, Authorization: Bearer and the witness. |
Lab-proven for this one principal. Header name was a column I set. |
| The same invoke for some other user's subject | confidentiality | Same lookup, their key, their later click. | Inferred. Not a second captured header. This is the UI:R. |
| Knocking the service over | availability | Not shown. | A:N. |
What stays closed
| Action | Result |
|---|---|
| Admin key following the new URL | Did not. Flag false, id-keyed secret gone, oracle did not contain the admin value. |
OAuth client secret when the flag is the string true |
Cleared on move. A separate question about a legacy plaintext column is not this page. |
| Cross-group name collision | 409. Not the oracle of this run. The check is still on the write path. |
Cross-group plant into another loom:group |
Closed in v1.10.1. This run stayed in demo. |
guarded_post to loopback, metadata, or 0.0.0.0/8 |
Still blocked. The lab address was private on purpose. |
| Secret-ARN plant, credential-provider delete through agent config, harness bind by display name | Closed in v1.10.1. Different bugs. |
The lab
python3 loom-mcp-userkey-follow-Abraxas-Labs.pyLoopback on the host, torn down after the line. One group, one principal, the witness stored as that principal's user key.
SUCCESS LOOM-190-MCP-USERKEY-FOLLOW user-key-followed endpoint-move LOOM-190-MCP-USERKEY-FOLLOW-WITNESSThat line is v1.9.0. On v1.10.3 the clear still stops at the admin key and the flagged OAuth secret. I re-read it. I did not rebuild it.
The fix
On an endpoint_url change, delete the per-user secrets for that server name, or refuse the move until they are gone. The admin-key block is the template: a credential issued for host A does not follow the row to host B. Name-keying can stay if deployed agents need the old path, but the clear has to happen before the new URL is saved, and a rename of the server is the same event. "The subject is in the path" tells you whose key it is. It does not tell you which host may see it.
References
- awslabs/loom at v1.10.3 (cc10ab0) and lab tag v1.9.0 (a305066).
user_api_key_secret_name.- Endpoint move: mcp.py.
- set_user_api_key.
- Invoke: invocations.py.
- Lab files: lab/.
Site: abraxaslabs.tech. Mail abraxas.null@proton.me.