Loom for AWS, issuer swap keeps the super-admin mapping
A security admin can change only the identity provider issuer. The super-admin mapping stays, login from the new issuer becomes g-admins-super, and the token proxy still posts the stored client secret to the new token endpoint. Labbed on v1.9.0. The same functions remain on v1.10.3, which I re-read and did not rebuild.
- Name
- Loom for AWS, issuer swap keeps the super-admin mapping
- Type
- 0-day analysis
- CVE
- n/a
- CVE Risk
- high
- Disclosure Status
- public
- Vendor
- Loom
- Affected
- Loom for AWS v1.9.0 (a3050669f5b4c83210a44ed4346f17e01641e82f), labbed, and v1.10.3 (cc10ab08573087fa12154f7c8181ffcd985fea88), source re-read. Caller is g-admins-security with a public HTTPS issuer they control.
- Published
- 16 Nov 2026
- Updated
- 16 Nov 2026
- Tags
- 0-day, loom, aws, cwe-863, cwe-522, identity-provider
The mapping stayed. The signer did not.
The lock they added is on the mapping table. The issuer URL is still a free field, and a free field is how a security admin becomes the super-admin.
The pack is abraxas/loom-idp-issuer-swap, the loopback client is loom-idp-issuer-swap-Abraxas-Labs.py, and I post as @abraxas_null.
Picture the person allowed to edit how the company signs in, and nothing else. They change the address of the login provider. The list that says which outside group counts as a super-admin stays put. The next token from the new address is accepted as that super-admin, and the secret Loom uses to finish a confidential login is posted to the new address. They already hold the security-admin job, and they need a public site that answers normal HTTPS discovery. No account means no attack. This does not run a program on the host.
Loom is the AWS Labs control plane for agents, groups, and a plug-in identity provider. I was in it because the bulletin had already closed three bugs and I wanted what the later patches left. CVE-2026-103956 is unauthenticated super-admin, fixed in v1.6.1. CVE-2026-103957 is OAuth discovery disclosure. CVE-2026-103958 is the MCP and A2A server-side request forgery. The bulletin is 2026-124. Not this page. v1.10.1 also closed four leftovers I am not reopening: a cross-group MCP plant, 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.
High 8.7, CVSS:3.1/AV:N/AC:L/PR:H/UI:N/S:C/C:H/I:H/A:N. CWE-863 and CWE-522. Remote, and the caller is already a security admin. I never showed the service going down. No CVE id here.
The container that printed SUCCESS was tag v1.9.0, commit a3050669f5b4c83210a44ed4346f17e01641e82f. I then opened the same functions on tag v1.10.3, commit cc10ab08573087fa12154f7c8181ffcd985fea88. The omit-and-swap is still there. I did not build v1.10.3 into a container for this issue.
g-admins-security holds security:read, security:write, and tagging:read. It does not hold admin:write. g-admins-super holds that scope and essentially every other one. t-admin by itself grants nothing. A normal deployment already has a row whose table maps some external group name onto t-admin and g-admins-super. That table is the policy. The issuer, and the keys discovery fills in from it, is who may assert the names in the policy. Swapping the signer without looking at the policy again is the bug.
The sink
update_identity_provider is the PUT. The comment says the guard is gated on "is not None" so no settable value skips it. Omitting the field is not a settable value. Both the assert and set_group_mappings run only when request.group_mappings is not None. Leave the key out and the old column stays, super-admin mapping included.
issuer_url is in the ordinary setattr loop. A changed value sets a flag and _run_discovery replaces jwks_uri, authorization_endpoint, and token_endpoint. client_secret_arn is not in that loop. Sending a new client secret would rewrite it. Sending only the issuer does not.
In plain terms, the spreadsheet that says "this outside group is the super-admin" is still on the wall. Only the lock on the front door was replaced, and the caller has the new key.
Discovery is not request forgery. fetch_discovery uses safe_get, public HTTPS, private addresses refused. The document's token endpoint has to be absolute HTTPS too. I needed a public issuer I controlled.
Login looks the token's group claim up in that table. An empty table grants nothing. A preserved super-admin line grants g-admins-super. The lookup assumes the person who can swap the issuer cannot keep a mapping they are forbidden to write.
The second sink is exchange_token, POST /api/auth/token, with no login required. If client_secret_arn is set, the server loads the secret and safe_posts it to whatever token_endpoint discovery just stored. On v1.10.3 the provider's response body is logged and is not copied back to the anonymous caller. The detail string is a fixed sentence. Smaller oracle. Not a fix. The secret has already left in the form body. MCP and A2A already delete a stored OAuth secret when their URL moves. This update does not.
to_dict is what the security admin can see first. List and PUT response include the mapping table, client_id, and has_client_secret. They do not include the secret string. You learn that a secret exists. You do not read it back through the API.
Wrong turns
I expected the empty-table pass-through from an older tag, where a missing map meant the claim values were the groups. Closed. An empty map plus a token claiming g-admins-super produced no super-admin group. The line records empty-map-super=no.
Writing the crown in directly is also closed. A PUT whose group_mappings include g-admins-super is 403. Creating a provider that names that group hits the same assert. The run created the privileged row as the bootstrap super-admin, which is how a deployment gets the row, then switched to the security caller.
The turn that worked was one field. issuer_url pointed at my issuer. No group_mappings key. Response 200, old mappings still in the body, has_client_secret still set, token_endpoint moved. An RS256 token from that issuer, claim equal to the external name already in the table, came back from GET /api/auth/me as t-admin,g-admins-super.
I checked whether discovery was the old SSRF. A private issuer is never fetched. On the v1.10.3 read I almost called the token proxy fixed, because the anonymous caller no longer receives the provider's reply. get_secret and safe_post above that log line are unchanged. The caller lost a reflection. The new token host did not lose the form field.
What an attacker can do
The caller starts as a security admin and ends holding admin:write. The lab wrote the witness through the site-settings route that had just returned 403, then read it back.
| Method | Class | What they get | Proof |
|---|---|---|---|
GET /api/settings/identity-providers |
recon | Mapping table, client_id, has_client_secret. Not the secret string. |
Same to_dict the issuer PUT returned. Lab saw the preserved map and the flag on that 200. |
PUT /api/settings/identity-providers/{id} with only issuer_url |
integrity | JWKS, authorize URL, and token URL replaced. Old mapping and client_secret_arn stay. |
Lab-proven. issuer-put=200. |
| RS256 login at the new issuer, claim equal to a preserved external key | integrity | Groups become t-admin,g-admins-super. |
Lab-proven. me-groups=t-admin,g-admins-super. |
PUT /api/settings/site/lab_witness, then GET /api/settings/site |
integrity | An admin:write call that was 403 returns 200 and the witness is stored. |
Lab-proven. admin-write=200, witness-written=yes. |
Unauthenticated POST /api/auth/token |
confidentiality | Loads client_secret_arn and posts client_secret to the new token_endpoint. |
Lab-proven. client-secret-match=yes. The match was the form body at my token endpoint, not a secret echoed to the client. |
| Taking the service down | availability | Not shown. | Not claimed. This is why the vector says A:N. |
What stays closed
| Action | Result |
|---|---|
Create a provider whose mappings name g-admins-super, as this caller |
403. Create asserts when the field is present. The lab's privileged row was created by the bootstrap super-admin. |
PUT that includes g-admins-super in group_mappings |
403. The run stops otherwise. |
| Login through an empty mapping table | No useful groups. empty-map-super=no. admin:write stays 403. |
| Discovery at a private or non-HTTPS issuer | Blocked by safe_get. Not CVE-2026-103958, and not a new SSRF. |
Provider error body on POST /api/auth/token in v1.10.3 |
Logged, not returned. The secret still moves. |
| Unauthenticated super-admin | CVE-2026-103956, fixed in v1.6.1. |
The lab
Host ports were 127.0.0.1:18600 and 127.0.0.1:18601, product and issuer. The stack came down after the line.
python3 loom-idp-issuer-swap-Abraxas-Labs.pySUCCESS LOOM-190-IDP-ISSUER-SWAP empty-map-super=no issuer-put=200 me-groups=t-admin,g-admins-super admin-write=200 witness-written=yes client-secret-match=yes LOOM-190-IDP-ISSUER-SWAP-WITNESSThe witness is LOOM-190-IDP-ISSUER-SWAP-WITNESS. The v1.10.3 read still shows the guard skipping an omitted field, discovery rewriting on issuer change, and the exchange attaching the stored client secret.
The fix
When issuer_url changes, re-check the mapping already stored, or refuse the update unless group_mappings is present and passes the assert. An omitted field must not mean "keep the privileged table, swap the signer." Clear client_secret_arn when discovery rewrites token_endpoint, the way an MCP endpoint move already clears the admin key, and do not POST the old secret until someone stores a new one for the new host. The fixed error string can stay. It is not the patch.
References
- Product: awslabs/loom, tags v1.9.0 and v1.10.3.
update_identity_providerand_run_discovery.exchange_token._map_external_groupsandg-admins-security.to_dict.- Lab tree: lab/.
I am on abraxaslabs.tech. Mail abraxas.null@proton.me.