
A Loom MCP admin in one group can move a connector to a new host. The admin key is deleted. Saved per-user API keys are not, and the next invoke sends them to that host. Same group only. v1.9.0 and v1.10.3. Medium 6.9.
An MCP administrator in a group can change a connector's address. Loom deletes the connector's own admin key when that happens. It does not delete the per-user keys saved under the connector's name. The next invoke sends those keys to the new host.
Same group only. Putting the connector into another group's tag, or renaming it onto a name that group already uses, is refused. The lab stored one user's key and invoked as that same user. A second person's key is the same name pattern and was not a second captured header.
Class: Secret follows a URL change
Reach: Remote. mcp:write on a connector the caller can already update. Another user's key is disclosed when that user invokes.
Loom v1.9.0 a3050669 and v1.10.3 cc10ab0
| ID | no CVE yet |
| Class | Per-user API key survives an endpoint move |
| Reach | Remote, same loom:group |
| CWE | CWE-522 |
| CVSS | Medium: 6.9 CVSS:3.1/AV:N/AC:L/PR:H/UI:R/S:C/C:H/I:L/A:N |
| Product | awslabs/loom |
| Affected | v1.9.0 a3050669f5b4c83210a44ed4346f17e01641e82f and v1.10.3 cc10ab08573087fa12154f7c8181ffcd985fea88 |
| Auth | g-admins-demo (mcp:write) in the connector's group |
| License | GNU Affero GPL v3.0 |
| Lab | loopback Docker on v1.9.0. v1.10.3 was a source check. |
UI:R is the other person invoking later. The container invoked once, as the admin who had just stored the witness, to show the header.
| Privilege | Method / sink | Impact | Lab-proven or inferred |
|---|---|---|---|
Move endpoint_url and clear the admin key | PUT /api/mcp/servers/{id} | Integrity. has_admin_api_key=false. The id-keyed admin secret is gone. | Lab-proven |
| Keep the per-user secret | user_api_key_secret_name(name, sub) is loom/mcp/{name}/api-key/{sub} | Confidentiality, still stored | Lab-proven. The witness was still in the secret store. |
| Send that secret to the new host | POST /api/agents/{id}/invoke with connector_ids | Confidentiality. Authorization: Bearer on the new host. | Lab-proven for the caller's own stored key |
| Send someone else's key the same way | the same lookup, their sub | Confidentiality | Inferred. Same function, not a second capture. |
Stays closed
| Axis | Method / sink | Why it is closed |
|---|---|---|
| Confidentiality | Admin key on the new host | Lab checked. The admin secret was absent and the admin value was not in the capture. |
| Integrity | Create or rename into another group's name | 409 |
| Confidentiality | Metadata and loopback fetches | guarded_post still refuses those. The lab oracle was an address the guard allows. RFC1918 is allowed on purpose. |
The admin secret was moved onto the server id so a rename could not aim it at another group. The per-user secret stayed on the display name, with a comment that the path is baked into agent config at deploy and that cross-group name collisions are refused. That closes the collision. It does not clear the key when the only thing that changed is the host.
The SUCCESS container was v1.9.0. user_api_key_secret_name on v1.10.3 is the same shape, and the endpoint move still does not delete that path.
python3 loom-mcp-userkey-follow-Abraxas-Labs.py
SUCCESS LOOM-190-MCP-USERKEY-FOLLOW user-key-followed endpoint-move LOOM-190-MCP-USERKEY-FOLLOW-WITNESS
run.sh checks out the public tag and tears the stack down.
On endpoint_url change, delete every loom/mcp/{name}/api-key/ secret for that connector, or key those secrets on the server id and stop writing the name into config. A comment that collisions are refused is not a clear.
cc10ab08573087fa12154f7c8181ffcd985fea88services/mcp.py user_api_key_secret_nameGNU Affero GPL v3.0. See LICENSE. Loopback lab only. No warranty.
Abraxas Labs [email protected]