Findings — auth: getting a token, and what it is

001_auth

Send the token request as application/x-www-form-urlencoded: application/json is 400 Invalid JSON body. client_credentials returns a 600s bearer, so ignore the refresh_token and re-auth.

• expires_in is 600 exactly, as documented. A long-running client must handle expiry, not assume one token per session.

• A refresh_token is issued but buys nothing here: re-authing is one call with credentials already in hand, so the client re-auths instead of storing refresh state.

• The token is a plain bearer string. One exchange against one site was measured, so whether a token or a script credential is accepted by a second site is untested. Settling it needs credentials on another site.

Content-Type. The endpoint matches the media type and ignores its parameters:

Two production clients strip the charset parameter before posting here, on the report that the endpoint rejects it. On the probed site it does not: every parameter above returns 200, including one naming a charset that does not exist. The 415 that names the accepted set comes from the media type in front of the parameter, so read that part of the header first. application/json is a different failure again: it is accepted as a media type and the form body is then parsed as JSON, giving 400 Invalid JSON body.

corpus/findings/001_auth.md

027_auth_permissions

The token endpoint accepts password and session_token. Impersonation is the OAuth2 scope sudo_as_login:, never a body field, and a lower level reads far fewer rows and as many fields.

• A script key is not the only way in. password and session_token are accepted grants: both fail on the credential, not on the grant name, while authorization_code, implicit and the device-code URI all fail on the grant name. A person with a Flow PT login can therefore reach the same REST API without an administrator issuing a script user.

• No route reports the caller. The bearer itself does: it is three dot-separated segments, and segment 2 is base64url JSON with a user claim of {type, id} plus auth_type and sudo_as_login. Decode it to read the claim; never verify it, and never log it.

• The permission model is readable as rows and opaque as rules. PermissionRuleSet has 10 fields, all read only, and none of them holds a rule. A client learns which set a user is in and nothing about what that set allows.

• Every measurement in this corpus was taken by a script user in the api_admin set, and a lower level reads far fewer rows and exactly as many fields. Impersonating an Artist on the probed site returned 8 projects of 22 and zero Versions, Shots, PublishedFiles and Notes, while /schema/<type>/fields returned the same counts at every level.

• A row count or a fill rate recorded anywhere in this corpus is an api_admin number; a field census is not level-dependent.

Grant types. Two failure shapes separate an accepted grant from a rejected one. Unsupported grant_type (code 103) means the name is not in the set. Any other error means the name was accepted and the credential was not.

Unlike a filter operator (probe 017), the rejection does not enumerate the accepted set: source is {} or null on all of them. The set above was found by probing names, so a grant this probe did not try may exist.

session_token is the grant the official launcher's flow ends at. GET /api/v1 reports authentication_app_session_launcher_enabled and unified_login_flow_enabled before any token exists, so a client can test a site for that path without credentials. Obtaining a session token needs the launcher and was not measured here.

Impersonation. sudo_as_login is an OAuth2 scope, not a body field. /spec.json declares it on both accepted grants as sudo_as_login:{user_login}, and only the scope form is read:

The target needs HumanUser.can_impersonate_this_user true and sg_status_list active. Each refusal names its own reason and all three land at the token endpoint, before any request the caller meant to make, so a client learns it cannot act as someone the moment it authenticates rather than part way through a job.

The flag is not settable over REST. A script in api_admin writing it is refused 400 code 104 The field is not editable for this user: [HumanUser.can_impersonate_this_user], so impersonation is something a site administrator grants in the web UI and a client can only read.

That last refusal was measured with the flag turned off by hand in the web UI, because no active account on the probed site had it set to false. The probe looks for one and prints the row only where the site has it, so a rerun elsewhere may show two refusals rather than three.

A sudo'd token does not change who the caller is: user stays the ApiUser, and sudo_as_login is added holding the login string the caller supplied. The token therefore reports nothing the caller did not already know, so a client wanting the human's id must look the login up.

Who am I. There is no me route under any of the seven shapes tried. The identity is in the token:

import base64, json
p = access_token.split(".")[1]
who = json.loads(base64.urlsafe_b64decode(p + "=" * (-len(p) % 4)))["user"]   # {"type": "ApiUser", "id": 298}

Then GET /entity/api_users/<id> or /entity/human_users/<id> returns permission_rule_set under relationships as a {id, name, type} hash, and projects as the multi_entity list that scopes the caller to a project subset. The ApiUser measured here reads every project with an empty list; a HumanUser with an empty list does not (entity_types/HumanUser). A script that cannot decode its own token can still find itself by filtering api_users on firstname is <the client_id it authenticated with>, because firstname is the script name; _search returned exactly one row for it.

What the permission model exposes. /entity/permissions, /entity/permission_rules and /entity/roles do not exist. PermissionRuleSet rows are the whole of it, keyed by entity_type: a plain type name is a set users are assigned to, and a dotted PermissionRuleSet.<Type> composite is the site's default set for that type. On the probed site there are 12 rows. Group exists as a separate 18-field type whose users multi_entity holds members; on the probed site all 4 groups are empty, and HumanUser.groups was empty on every row.

On the probed site the 24 HumanUser rows split 12 Admin, 6 Artist, 3 Manager, 3 Vendor, and all 16 ApiUser rows are API Admin.

What a level changes. The experiment this finding once called for, a second caller at a lower level diffed row for row, needs no second credential: impersonation supplies it. Measured with one script key, acting as itself, as an Admin and as an Artist:

rows projects          script:22   Admin:22   Artist:8
rows human_users       script:25   Admin:25   Artist:25
rows api_users         script:16   Admin:16   Artist:400 "Entity of type ApiUser can not be
                                                          accessed by this user. Rule: Artist"
rows versions          script:200  Admin:200  Artist:0
rows shots             script:200  Admin:200  Artist:0
rows published_files   script:200  Admin:200  Artist:0
rows notes             script:200  Admin:200  Artist:0
schema fields Version  script:71   Admin:71   Artist:71
GET /license_info      script:200  Admin:200  Artist:401 "Must sudo as Administrator to query
                                                          license information"
GET /me                script:404  Admin:404  Artist:404

Rows collapse, fields do not, and a refusal is a 400 naming the rule or a 401 naming the level rather than an empty list. /me stays absent for a human caller too, so that finding is a property of the API and not of the script user who first measured it.

Unmeasured. Which permission rules produce those numbers: PermissionRuleSet exposes no rule and the site's own settings are not on the REST surface, so only the effect is measurable, never the cause. Conditional permissions are therefore invisible here. Whether password or session_token yields a usable REST token at all, and what auth_type then reads, is still untested: both were sent deliberately fake credentials. Only levels the probed site has active can be compared, and a site whose lower-level accounts are all disabled measures nothing.

corpus/findings/027_auth_permissions.md

052_app_session_launcher

Post appName and machineId, open url in a browser, PUT the id until approved. The sessionToken spends at grant_type=session_token as that person, and every mint renews the session. [partial]

not measured: a person clicking deny, and a session left idle past the site's expiry window; both wait on time and on the site, not on the probe

• A person reaches the REST API with no script key and no password: two unauthenticated calls, a browser, and a click. The bearer is a HumanUser, so created_by and Version.user are the person without sudo_as_login and without an administrator granting can_impersonate_this_user (probe 027).

• The session token is the credential; the bearer is disposable. Hold the session token, mint a 600s bearer whenever one is needed, and each mint moves the session's expiry to now plus the site's window (one day on the probed site), written at most once every five minutes.

• A token spent at least once per window never expires; one left idle past it does, and the token endpoint then refuses it.

• A pending request lives about five minutes and a handed-out one is gone at once. Poll from one place, keep sessionToken from the one response that holds it, and when the poll turns 404 issue a new request and show the person the new url rather than tell them what went wrong: forgotten, denied and mistyped all read {"message":"Not Found"}.

• /internal_api is the web app's own surface, found by reading the site's session checker script. Every call on it takes the session token as the _session_id cookie, answers errors as {"message": ...} rather than errors[], and is versioned by nobody. Read the site's GET /internal_api/session for the expiry rather than assume the preference.

corpus/findings/052_app_session_launcher.md