Loom for AWS, agent logs skip the session owner
A t-user in an agent's group, with session:read and agent:read, is refused on another user's session and allowed on that session's logs, traces, and evaluation exchange. Same group only. The session-body IDOR is closed. Labbed on v1.9.0. The owner check is still missing on v1.10.3, which I re-read and did not rebuild.
- Name
- Loom for AWS, agent logs skip the session owner
- Type
- 0-day analysis
- CVE
- n/a
- CVE Risk
- medium
- Disclosure Status
- public
- Vendor
- Loom
- Affected
- Loom for AWS v1.9.0 (a3050669f5b4c83210a44ed4346f17e01641e82f), labbed, and v1.10.3 (cc10ab08573087fa12154f7c8181ffcd985fea88), source re-read. Caller is a t-user in the agent group. Read of prompts and answers.
- Published
- 16 Nov 2026
- Updated
- 16 Nov 2026
- Tags
- 0-day, loom, aws, cwe-639, logs
The session said no. The logs said the prompt.
Alice can open her own chat. Bob, on the same agent and in the same group, is refused that chat. The log route, given the same agent and the same conversation, returns 200 and includes what she typed.
Pack: abraxas/loom-logs-owner. Loopback client: loom-logs-owner-Abraxas-Labs.py. I post as @abraxas_null.
Two people on one team share an agent. The session page knows who owns a conversation. The owner can read it. A teammate cannot. Logs, the trace, and the evaluation view that returns what was asked and what was answered do not ask that question. They ask whether the caller is in the agent's group and holds a read scope. Bob is, so he gets the prompt and the answer. Charlie, on a different team, is still refused. Nothing here edits the agent or runs code. It reads text that was already written down. A trace, if the word is muddy, is the same chat copied for the operator console, spans included.
I was in Loom after the public set was already named. CVE-2026-103956 is unauthenticated super-admin, fixed in v1.6.1. CVE-2026-103957 is OAuth discovery. CVE-2026-103958 is the MCP and A2A SSRF. Bulletin 2026-124. Not this. v1.10.1 closed a cross-group MCP plant, planting a secret ARN in agent config, deleting a credential provider through that config, and harness invoke binding an MCP server by display name. Closed, and not a log read.
Medium 6.5, CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:H/I:N/A:N, CWE-639. Remote HTTP. PR:L is a normal user in the group, not the open internet and not a super-admin. Scope stays unchanged because the text is already that group's conversation store. I did not write anything as Bob, and I did not show an outage, so integrity and availability are N. No CVE id on this post.
The SUCCESS line came off a container of tag v1.9.0, commit a3050669f5b4c83210a44ed4346f17e01641e82f. I then read tag v1.10.3, commit cc10ab08573087fa12154f7c8181ffcd985fea88. The owner check is still absent from the log, trace, and exchange handlers. I did not rebuild on v1.10.3. The statuses below are the v1.9.0 run. The line numbers are the later tag.
Bob is t-user plus g-users-demo. Alice is the same kind of user, and she owns the session. Charlie is t-user plus g-users-test. The agent is tagged demo. Shipped user groups that can see an agent carry both agent:read and session:read. The scope comment says session:read is held alongside agent:read by everyone who has it today, and that splitting the scope is not a privilege change. The point of the split was that conversation text is more sensitive than "this agent exists." That sentence is true for the session route. It is not true for the routes that return the same sentences from the logs.
t-admin may read other people's conversations in the agent's group. assert_session_readable returns immediately when that group is present. Bob is not a t-admin. This post is not "admins can read logs."
Two checks, and one of them is missing
get_session loads the row through get_session_or_404, which checks the agent's group and then calls assert_session_readable. The owner may read the session. An admin for the agent's group may read it. Anyone else, if session.user_id is set and is not the caller's username, gets 403. Alice's GET was 200. Bob's GET of the same URL was 403, and the witness was not in the error. That is the closed bug. The old session-body IDOR, where a UUID skipped both checks, stays closed.
_resolve_agent_and_log_group is what the log routes call instead. It exists so the group check happens once. get_agent_or_404 closes the older hole, another group's logs by id. The helper validates the qualifier and derives the log group from the runtime id. It never calls assert_session_readable. It never sees a session id.
GET /api/agents/{id}/logs requires session:read and then that helper. The lab passed the stream name and got the events, witness included. GET /api/agents/{id}/sessions/{sid}/logs is the same helper, filtered by session id. Same missing call. I am not splitting it. The run's 200 is the logs GET.
GET /api/agents/{id}/traces/{trace_id} is the same miss: session:read, get_agent_or_404, span bodies back, no owner check. Bob's trace GET was 200, witness in the body.
The exchange route looked like a second bug and is not one. GET /api/agents/{id}/evaluations/traces/{trace_id}/exchange asks only for agent:read, then get_agent_or_404, and returns prompt and answer. A group with agent:read and not session:read would be a narrower door. Shipped groups are not built that way. Bob already has session:read. One bug.
A session is the chat a person sees. Logs are that chat copied for operators. The owner check landed on the first copy and not the second.
Wrong turns
The session route fooled me. Bob's 403, witness absent from that body, and I nearly wrote the class off. The line keeps bob-session=403 so the closed session IDOR stays closed in the oracle.
The log helper's comment talks only about groups. The function never receives a session. Bob's logs GET was 200. Charlie, in g-users-test, was 403, witness absent. That is the cross-group limit, and it is in the line.
I still fetched the exchange, because "same class" is a claim you measure. Every shipped group that can call it can already call the log route. Vended logs are a different, closed hole: get_vended_logs allowlists the agent's own runtime and memories, so a caller-named group is not this bug and is not in the SUCCESS line. Labbing Bob as t-admin would have been the product working. The run checks he is t-user in g-users-demo, and not t-admin.
What an attacker can do
Bob already has session:read and agent:read because he is a user in the agent's group. He does not have the session.
| Method | Class | What they get | Proof |
|---|---|---|---|
GET /api/agents/{id}/sessions/{sid} as Alice |
recon, control | Her session, witness included. | Lab-proven. alice-session=200. |
GET /api/agents/{id}/sessions/{sid} as Bob |
confidentiality, refused | 403. Witness not in the body. | Lab-proven. bob-session=403. The closed session IDOR, listed so it is not confused with the open route. |
GET /api/agents/{id}/logs as Bob |
confidentiality | Log events, witness in the body. No owner check. | Lab-proven. logs=200, logs-has-witness=yes. |
GET /api/agents/{id}/sessions/{sid}/logs |
confidentiality | Same helper, filtered to the session id. | Same missing call. Inferred from the handler, not a second SUCCESS field. |
GET /api/agents/{id}/traces/{trace_id} as Bob |
confidentiality | Span bodies, witness present. | Lab-proven. traces=200, traces-has-witness=yes. Same class. |
GET /api/agents/{id}/evaluations/traces/{trace_id}/exchange as Bob |
confidentiality | Prompt and answer. agent:read plus the group check. Shipped groups with agent:read also have session:read. |
Lab-proven. exchange=200, exchange-has-witness=yes. Same class, not a second bug. |
| Writes, or stopping the service | integrity, availability | Not shown. | I:N, A:N. |
No cross-group row in that table on purpose.
What stays closed
| Action | Result |
|---|---|
| Session body for a non-owner | 403 via assert_session_readable. Witness does not leak in the error. |
Charlie, g-users-test, reading the demo agent's logs |
403. charlie-logs=403. |
t-admin reading another user's in-group transcript |
Allowed. Intentional. |
| Caller-named vended log group outside the agent's own runtime and memories | Allowlisted. Closed older hole. Not in this SUCCESS line. |
| Unauthenticated read | No. Bob is a user. |
| The four v1.10.1 leftovers, and the three public CVEs | Closed. Not this page. |
The lab
The product listened on 127.0.0.1:18630. The witness lived in Alice's session and in the log and trace text. The stack came down after the line.
python3 loom-logs-owner-Abraxas-Labs.pySUCCESS LOOM-190-LOGS-OWNER alice-session=200 bob-session=403 logs=200 logs-has-witness=yes charlie-logs=403 exchange=200 exchange-has-witness=yes traces=200 traces-has-witness=yes LOOM-190-LOGS-OWNER-WITNESSThat line is v1.9.0. I did not boot v1.10.3 to re-print it.
The fix
Call the owner rule after the group check, everywhere the response is conversation text. _resolve_agent_and_log_group cannot do it alone, because it has no session, but the session-log route does, and the trace and exchange handlers are already looking at events tied to one. The predicate exists. assert_session_readable is owner, or t-admin for that agent. Use it before the events are returned. Changing the exchange route from agent:read to session:read would match the scope comment and would not stop Bob. He has both. The bug is the missing owner check.
References
- awslabs/loom, lab tag v1.9.0 (a305066), read tag v1.10.3 (cc10ab0).
_resolve_agent_and_log_groupandget_agent_logs.assert_session_readableandget_session.get_trace_detail.get_evaluated_exchange.- Scope table: auth.py.
- Lab tree: lab/.