Returning objects in a collection vs. IDs

[silverpill@mitra.social]

@julian How should it be amended? Just add a sentence saying that some collection items might be full objects instead of IDs?

You can trust the objects owned by the server you're requesting data from. (this is fe34)

It's complicated. Those objects are embedded within another object (a collection), and FEP-fe34 currently doesn't recommend trusting embedded objects, except for a few special cases:

An object is the object of a Create activity, and it has the same origin and the same owner as the activity.
An embedded object is identified as a fragment of the wrapping object.
An embedded object is anonymous (doesn't have an ID).

fep/fep/fe34/fep-fe34.md at b7dddd105b23762d4c59b4ffa7e43ca66af0939b

fep - Fediverse Enhancement Proposals

Codeberg.org (codeberg.org)

This is because ActivityPub Clients can publish arbitrary activities, and a malicious actor may embed an object that is attributed to another actor on the same domain. The server hosting the malicious actor can't realistically verify every embedded object, so if a remote server trusts same-origin embedded objects unconditionally, the malicious actor might succeed in impersonating another actor.

But maybe collection items should be added to the list of exceptions, because collections are almost always server-managed? I need to think about this.

@grishka

[grishka@mastodon.social]

@silverpill @julian @technical-discussion
> The server hosting the malicious actor can't realistically verify every embedded object

How so? It's a responsibility of that server to authenticate its actors, and restrict them from impersonating other actors on the same server, C2S or not. I don't quite understand the threat model here. In mine, clients are always assumed malicious.

[julian@activitypub.space]

grishka@mastodon.social silverpill@mitra.social yes I had the same question.. I assumed this would only occur in a C2S context but even then the server is an intermediary and can verify.

[silverpill@mitra.social]

@julian @grishka

Here's en example:

{
  "type": "Create",
  "actor": "https://social.example/alice"
  "object": {
    "type": "Note",
    "attributedTo": "https://social.example/alice",
    "attachment": {
      "type": "Note",
      "attributedTo": "https://social.example/bob"
    }
  }
}

It contains an embedded Note that is attributed to another actor. There are many possible ways to embed an object, and malicious embedding could be difficult to detect for the origin server.

[trwnh@mastodon.social]

@julian @grishka @silverpill

> messy... mixing types

if mixing types is a concern you can serve ids in object form: {"id": "foo"}

[grishka@mastodon.social]

@silverpill @julian @technical-discussion ideally you would traverse everything and replace anything you don't understand with "id" references.

But anyway, I feel like we're getting too carried away about a very niche aspect of the whole thing. Almost like on that SocialHub forum.

In my own AP extensions I always just act like C2S doesn't exist because I've never seen it used in practice, and it's wildly impractical to use anyway.

[trwnh@mastodon.social]

@silverpill @julian @technical-discussion @grishka this wouldn't be "malicious", it would be "unverified". anyone can say anything about anything. it'd be like if i said herman melville wrote moby dick, and you didn't trust me until you verified with herman melville himself that he wrote moby dick. it's just a statement that you can either accept or reject. maybe you don't trust alice to make claims about bob, or maybe you trust social.example to make claims on social.example

[trwnh@mastodon.social]

@grishka @silverpill @julian @technical-discussion what does exist is publishing an AS2 resource on a web server, which you can do with any "api" you want -- ftp, rsync, ssh, anything that can get a file in a folder served by nginx with the appropriate headers. "c2s" is literally just an endpoint that you can POST some json to and it will publish that json and POST it to others. you can replace it with curl.

[grishka@mastodon.social]

@trwnh @silverpill @julian @technical-discussion if I remember the spec correctly, that endpoint does have to do at least *some* interpretation of the json — for example, the spec explicitly says that Block activities that come into the inbox "MUST NOT" be exposed in the client's view of the inbox, but instead interpreted and acted upon by the server itself

[trwnh@mastodon.social]

@grishka @silverpill @julian @technical-discussion i think you're referring to how Block should not be sent to an inbox if posted to an outbox? but this is just a rule for outboxes. the user/agent can manually send whatever notifications they want. the user/agent can also follow those rules and just store the relevant json on any properly configured web server.

[trwnh@mastodon.social]

@grishka @silverpill @julian @technical-discussion point being, you can completely upend a lot of fedi applications' security models with just nginx and curl and subdirectories.

[grishka@mastodon.social]

@trwnh @silverpill @julian @technical-discussion I still wouldn't call that "upending". The actor objects — that reside on that server — contain public keys, which the entire security model of the entire fediverse relies on. So it follows, then, that since a server can serve any arbitrary public key for any of its actors, and the rest of the fediverse will unquestionably trust it, it can also be trusted with any other objects on the same domain.

[trwnh@mastodon.social]

@grishka @silverpill @julian @technical-discussion the key can be on a different keyserver

[grishka@mastodon.social]

@trwnh @silverpill @julian @technical-discussion it can in theory but no software currently in existence supports that

[grishka@mastodon.social]

@trwnh @silverpill @julian @technical-discussion but even if, the actor object is still the ultimate source of truth as it's the one which contains the key ID

[trwnh@mastodon.social]

@grishka @silverpill @julian @technical-discussion true but we are talking about mistakes. like taking the origin of the keyId on the http sig instead of following the link to the owner/controller. such assumptions might be true for monoliths, but not everything is a monolith

[thisismissem@activitypub.space]

Web Annotations has a neat solution here which is prefers representation, which allows the requestor to ask for minimal or full representation (i.e., just the IDs or the full representations)

You'd still likely want to gate that on authorization

[silverpill@mitra.social]

@grishka I am developing a client application where this is a real concern.
But I agree that in general, originating servers are responsible for verification of client data. This part of FEP-fe34 will likely be revised in the future.

[mariusor@metalhead.club]

> and it's wildly impractical to use anyway.

@grishka shush, you! 🤫

@silverpill @julian @technical-discussion

[mariusor@metalhead.club]

@silverpill do you mean that the "malicious" attachment is not a facsimile of an actual note produced by that actor, but a forgery?

In these cases, I'll agree with
@grishka that some validation based on the ID should be necessary.

For embedded object attachments on the other hand (like mastodon produces), probably the validation needs to check that attributedTo corresponds to the one of the parent object or missing.

Interesting corner case.

@technical-discussion