Loom, a registry editor rewrites a skill another group already attached
A registry admin who is not a super-admin rewrites a skill. The next full deploy by someone who can update another group's agent folds that markdown into the system prompt. The writer cannot open or invoke that agent.
- Name
- Loom, a registry editor rewrites a skill another group already attached
- Type
- 0-day analysis
- CVE
- n/a
- CVE Risk
- medium
- Disclosure Status
- public
- Vendor
- Loom
- Affected
- Loom for AWS. Container SUCCESS on v1.9.0 (a3050669f5b4c83210a44ed4346f17e01641e82f). The same functions are still present on v1.10.3 (cc10ab08573087fa12154f7c8181ffcd985fea88). This was not re-run in a container on v1.10.3. Needs registry:write (g-admins-demo or g-admins-registry), not g-admins-super. The fold waits for a later full deploy or update that still names the skill.
- Published
- 16 Nov 2026
- Updated
- 16 Nov 2026
- Tags
- 0-day, loom, aws, cwe-863, cwe-74, registry, authenticated
The check steps aside when the record is a skill
I went looking for the group check that was added so one catalog admin could not stamp APPROVED on another group's row. The docstring told me where that check does not apply. Skill records have no Loom row. The function returns, and registry:write is enough to replace the markdown.
The notes and the loopback client are on GitHub: abraxas/loom-skill-plant, loom-skill-plant-Abraxas-Labs.py. @abraxas_null. The product is awslabs/loom.
A coworker who can edit the shared skill catalog, and who is not a super-admin, can replace the text of a skill another team already attached. Nothing about that other team's agent changes until someone allowed to update it does a full deploy, or a full update that still names the skill. Loom then copies the new text into the agent's instructions. The editor cannot open the agent, cannot run it, and cannot approve the other team's linked catalog rows. Later replies that might follow those instructions are a guess, which is why confidentiality on the score is none.
No CVE. CWE-863 and CWE-74. Medium 6.2 (CVSS:3.1/AV:N/AC:L/PR:H/UI:R/S:C/C:N/I:H/A:N). Remote. PR:H is registry:write. UI:R is that later full deploy. S:C because the integrity lands in another group's agent. C:N / I:H / A:N. Confidence about 86 percent on the fold I watched. Not code execution. The thing that moves is instruction text.
How I got here
Loom sits in front of Bedrock AgentCore. g-admins-demo and g-admins-registry carry registry:write. Neither is g-admins-super, and both are refused on another group's agent. The skill write is allowed. The read of the victim is not.
I pinned v1.9.0, a3050669f5b4c83210a44ed4346f17e01641e82f, after the public bulletin. I was not going to refile CVE-2026-103956 (unauthenticated super-admin, fixed in v1.6.1), CVE-2026-103957 (OAuth disclosure, fixed in v1.7.0), or CVE-2026-103958 (MCP and A2A request forgery, fixed in v1.7.0). Unauthenticated websocket invoke is closed as well: first frame JWT, close 1008. This post is the skill exception the group check documents.
MCP servers, A2A agents, and agents have a registry_record_id. Approving one of those without the matching loom:group is a 403, and the lab printed that. Skills live only in the registry service. Loom never inserts a row for them.
The sink
_owned_resource_or_403 at registry.py line 126 looks up a Loom row and, on a hit, checks loom:group. The docstring says it returns None when no row is linked, and that skill records stay governed by registry:write. The body matches the docstring: _find_resource_by_record_id misses, and the function returns None.
PUT /api/registry/records/{record_id} calls that helper, then looks for a linked MCP server, a linked A2A agent, and a linked agent. When all three miss and the registry type is SKILL, the handler requires skill_name, skill_description, skill_license, skill_version, and skill_md, builds descriptors from them, and writes. The markdown is the caller's. Nothing compares it to the agents that already point at the record.
The fold is _get_attached_skill_prompt_text at agents.py line 786. It loads every enabled skill integration and fetches the registry record live, not from a cache. Status must be APPROVED. There is no loom:group comparison. Approved markdown is appended. _build_system_prompt at line 902 puts that string on the end of the system prompt. The heading at line 869 is ## Attached Skills.
A skill is a markdown document in the shared catalog. Attaching it stores a record id. The paste happens on a full deploy or a full update that still names that id.
POST /api/agents/{id}/redeploy resends the stored config and does not refetch. The full routes that still name the skill_id do. The form shows a name and a description, not the body, and the dependents count is the editor's own group only.
The fold demands APPROVED, and the same registry:write principal can POST /api/registry/records/{record_id}/approve. The helper returns None again. Put the witness in skill_md, approve, wait for the other group's full update. The stored prompt contains the heading and the witness.
The turn I got wrong
I treated it as an IDOR on the agent. GET returned 403. Invoke is the same group check, and I will not invent a cross-group invoke the lab did not print. The writer never talks to the agent. The integrity is the catalog object the victim already attached by id.
Then I approved a linked victim MCP row and a linked victim agent row. Both 403. The helper works when it finds a row. Skills return None on purpose. Not a super-admin bug.
Quick redeploy was the third miss. It keeps the previous prompt. UI:R is the wait for someone who can update that agent to run a full deploy or a full update that still names the skill.
What that account can do
The lab caller was t-admin plus g-admins-demo. That is registry:write, not g-admins-super. g-admins-registry holds the same registry scope and is the other principal this applies to. I did not log the lab in as that second group.
| Method | What it grants | Axis | Lab or inferred |
|---|---|---|---|
Unauthenticated GET of the caller |
Nothing. 401. | Closed | Lab-proven. unauth-me=401 |
Session as t-admin, g-admins-demo |
registry:write on skill records. Not super-admin. |
Recon of own groups | Lab-proven. me-groups=t-admin,g-admins-demo |
GET the other group's agent |
Refused. | Confidentiality closed | Lab-proven. agent-get=403 |
| Invoke that agent | Refused by the same group check. The writer never runs it. | Closed | Inferred from the 403 and the check. Not a separate oracle line. |
PUT /api/registry/records/{record_id} with caller-supplied skill_md |
Replaces the SKILL body. No Loom row, so no loom:group check. |
Integrity | Lab-proven. skill-put=200 |
POST /api/registry/records/{record_id}/approve on that skill |
Status becomes APPROVED, which is what the fold requires. | Integrity | Lab-proven. skill-reapprove=200 |
| Approve a linked victim MCP record | Refused. | Integrity closed | Lab-proven. mcp-approve=403 |
| Approve a linked victim agent record | Refused. | Integrity closed | Lab-proven. agent-approve=403 |
Full deploy or full update that still names the skill_id |
Live fetch, APPROVED only, append under ## Attached Skills. |
Integrity of the system prompt. Scope change. | Lab-proven. fold-witness=yes |
POST /api/agents/{id}/redeploy (quick) |
Resends the stored prompt. The new body is not folded. | Integrity does not move | Source. The route does not call the fetch. |
| Whatever the agent says afterwards | Would follow the planted instructions, including tools already bound to that agent. | Confidentiality of those replies | Inferred. That is why the vector is C:N. |
| Delete or crash the agent | Nothing on this path. | Availability none | Not claimed. A:N. |
I showed the witness inside the prompt deploy stored. I did not show a later answer, so I am not scoring a data leak. A pause rule already configured on that agent still applies to its tools. This changes the instructions. It does not turn the pause off.
What stays shut
| Surface | Why it stays closed |
|---|---|
Super-admin, g-admins-super |
Not required. The lab caller is a registry admin. |
GET or invoke of the other group's agent |
loom:group. Lab agent-get=403. |
| Approve or rewrite a linked MCP, A2A, or agent record | _owned_resource_or_403 finds the row and checks the group. Lab 403 on MCP and on agent. |
Create or update another group's resource by sneaking a resource_id |
The linked-row path 403s before the registry call. |
| Dependents list as a map of victims | Filtered to the caller's tag. The editor does not see the other group's agents. |
| Fold of a skill that is not APPROVED | The fetch skips it. |
| Quick redeploy as a trigger | Keeps the previous prompt. |
| Unauthenticated websocket invoke | First frame JWT, close 1008. Not this bug. |
| CVE-2026-103956, CVE-2026-103957, CVE-2026-103958 | Already public, already fixed on older tags. Not refiled. |
| Code execution | Not this. The artifact is markdown in a system prompt. |
The lab
The client talks to loopback. It plays both sides, the registry editor and then the update a victim-side agent:write would perform, which is why UI:R is on the vector even though one script prints the line.
python3 loom-skill-plant-Abraxas-Labs.pySUCCESS LOOM-190-SKILL-PLANT unauth-me=401 me-groups=t-admin,g-admins-demo agent-get=403 skill-put=200 skill-reapprove=200 mcp-approve=403 agent-approve=403 fold-witness=yes LOOM-190-SKILL-PLANT-WITNESSThe container that printed that line was Loom v1.9.0, commit a3050669f5b4c83210a44ed4346f17e01641e82f. I read the same functions on v1.10.3, commit cc10ab08573087fa12154f7c8181ffcd985fea88. _owned_resource_or_403 still returns None for a skill, and _get_attached_skill_prompt_text still appends approved markdown with no group comparison. I did not bring v1.10.3 up in a container and re-run this, or the two other leftovers from the same pass.
The change I would make
Stop treating "no Loom row" as permission for any registry:write principal to replace a body another group already consumes. Persist an owner group on the skill and run the same loom:group check used for MCP and agent records, including on approve. If the catalog is meant to be shared, pin the body or a hash at attach time and do not fold a different body until the group that owns the agent accepts it. Show that body on the deploy form. The live fetch is a fair way to honor an un-approval. It is a bad way to honor a silent edit.
References
- awslabs/loom v1.9.0
a3050669f5b4c83210a44ed4346f17e01641e82f, v1.10.3cc10ab08573087fa12154f7c8181ffcd985fea88 _owned_resource_or_403andupdate_record_get_attached_skill_prompt_textand_build_system_prompt- CWE-863, CWE-74
- CVSS 3.1 calculator
abraxaslabs.tech · github.com/abraxas · abraxas.null@proton.me