"id" is used for the canonical link to the AS2 representation of an object.
"url" is typically used for the canonical link to the HTTP representation.
It is what we use, for example, when following the "external source" link
in the frontend. However, it's not the link we include in the post contents
for quote posts.
Using URL instead means we include a more user-friendly URL for Mastodon,
and a working (in the browser) URL for Threads
previously we would uncritically take data and format it into
tags for static-fe and the like - however, instances can be
configured to disallow unauthenticated access to these resources.
this means that OG tags as a vector for information leakage.
_technically_ this should only occur if you have both
restrict_unauthenticated *AND* you run static-fe, which makes no
sense since static-fe is for unauthenticated people in particular,
but hey ho.
This protects us from falling for obvious spoofs as from the current
upload exploit (unfortunately we can’t reasonably do anything about
spoofs with exact matches as was possible via emoji and proxy).
Such objects being invalid is supported by the spec, sepcifically
sections 3.1 and 3.2: https://www.w3.org/TR/activitypub/#obj-id
Anonymous objects are not relevant here (they can only exists within
parent objects iiuc) and neither is client-to-server or transient objects
(as those cannot be fetched in the first place).
This leaves us with the requirement for `id` to (a) exist and
(b) be a publicly dereferencable URI from the originating server.
This alone does not yet demand strict equivalence, but the spec then
further explains objects ought to be fetchable _via their ID_.
Meaning an object not retrievable via its ID, is invalid.
This reading is supported by the fact, e.g. GoToSocial (recently) and
Mastodon (for 6+ years) do already implement such strict ID checks,
additionally proving this doesn’t cause federation issues in practice.
However, apart from canonical IDs there can also be additional display
URLs. *omas first redirect those to their canonical location, but *keys
and Mastodon directly serve the AP representation without redirects.
Mastodon and GTS deal with this in two different ways,
but both constitute an effective countermeasure:
- Mastodon:
Unless it already is a known AP id, two fetches occur.
The first fetch just reads the `id` property and then refetches from
the id. The last fetch requires the returned id to exactly match the
URL the content was fetched from. (This can be optimised by skipping
the second fetch if it already matches)
05eda8d193/app/helpers/jsonld_helper.rb (L168)63f0979799
- GTS:
Only does a single fetch and then checks if _either_ the id
_or_ url property (which can be an object) match the original fetch
URL. This relies on implementations always including their display URL
as "url" if differing from the id. For actors this is true for all
investigated implementations, for posts only Mastodon includes an
"url", but it is also the only one with a differing display URL.
2bafd7daf5 (diff-943bbb02c8ac74ac5dc5d20807e561dcdfaebdc3b62b10730f643a20ac23c24fR222)
Albeit Mastodon’s refetch offers higher compatibility with theoretical
implmentations using either multiple different display URL or not
denoting any of them as "url" at all, for now we chose to adopt a
GTS-like refetch-free approach to avoid additional implementation
concerns wrt to whether redirects should be allowed when fetching a
canonical AP id and potential for accidentally loosening some checks
(e.g. cross-domain refetches) for one of the fetches.
This may be reconsidered in the future.
Since we always followed redirects (and until recently allowed fuzzy id
matches), the ap_id of the received object might differ from the iniital
fetch url. This lead to us mistakenly trying to insert a new user with
the same nickname, ap_id, etc as an existing user (which will fail due
to uniqueness constraints) instead of updating the existing one.
Since we reject cross-domain redirects, this doesn’t yet
make a difference, but it’s requried for stricter checking
subsequent commits will introduce.
To make sure (and in case we ever decide to reallow
cross-domain redirects) also use the final location
for containment and reachability checks.
In order to properly process incoming notes we need
to be able to map the key id back to an actor.
Also, check collections actually belong to the same server.
Key ids of Hubzilla and Bridgy samples were updated to what
modern versions of those output. If anything still uses the
old format, we would not be able to verify their posts anyway.
If it’s not already in the database,
it must be counterfeit (or just not exists at all)
Changed test URLs were only ever used from "local: false" users anyway.
This brings it in line with its name and closes an,
in practice harmless, verification hole.
This was/is the only user of contain_origin making it
safe to change the behaviour on actor-less objects.
Until now refetched objects did not ensure the new actor matches the
domain of the object. We refetch polls occasionally to retrieve
up-to-date vote counts. A malicious AP server could have switched out
the poll after initial posting with a completely different post
attribute to an actor from another server.
While we indeed fell for this spoof before the commit,
it fortunately seems to have had no ill effect in practice,
since the asociated Create activity is not changed. When exposing the
actor via our REST API, we read this info from the activity not the
object.
This at first thought still keeps one avenue for exploit open though:
the updated actor can be from our own domain and a third server be
instructed to fetch the object from us. However this is foiled by an
id mismatch. By necessity of being fetchable and our longstanding
same-domain check, the id must still be from the attacker’s server.
Even the most barebone authenticity check is able to sus this out.
Such redirects on AP queries seem most likely to be a spoofing attempt.
If the object is legit, the id should match the final domain anyway and
users can directly use the canonical URL.
The lack of such a check (and use of the initially queried domain’s
authority instead of the final domain) was enabling the current exploit
to even affect instances which already migrated away from a same-domain
upload/proxy setup in the past, but retained a redirect to not break old
attachments.
(In theory this redirect could, with some effort, have been limited to
only old files, but common guides employed a catch-all redirect, which
allows even future uploads to be reachable via an initial query to the
main domain)
Same-domain redirects are valid and also used by ourselves,
e.g. for redirecting /notice/XXX to /objects/YYY.
No new path traversal attacks are known. But given the many entrypoints
and code flow complexity inside pack.ex, it unfortunately seems
possible a future refactor or addition might reintroduce one.
Furthermore, some old packs might still contain traversing path entries
which could trigger undesireable actions on rename or delete.
To ensure this can never happen, assert safety during path construction.
Path.safe_relative was introduced in Elixir 1.14, but
fortunately, we already require at least 1.14 anyway.
To save on bandwith and avoid OOMs with large files.
Ofc, this relies on the remote server
(a) sending a content-length header and
(b) being honest about the size.
Common fedi servers seem to provide the header and (b) at least raises
the required privilege of an malicious actor to a server infrastructure
admin of an explicitly allowed host.
A more complete defense which still works when faced with
a malicious server requires changes in upstream Finch;
see https://github.com/sneako/finch/issues/224
Certain attacks rely on predictable paths for their payloads.
If we weren’t so overly lax in our (id, URL) check, the current
counterfeit activity exploit would be one of those.
It seems plausible for future attacks to hinge on
or being made easier by predictable paths too.
In general, letting remote actors place arbitrary data at
a path within our domain of their choosing (sans prefix)
just doesn’t seem like a good idea.
Using fully random filenames would have worked as well, but this
is less friendly for admins checking emoji dirs.
The generated suffix should still be more than enough;
an attacker needs on average 140 trillion attempts to
correctly guess the final path.
This will decouple filenames from shortcodes and
allow more image formats to work instead of only
those included in the auto-load glob. (Albeit we
still saved other formats to disk, wasting space)
Furthermore, this will allow us to make
final URL paths infeasible to predict.
Since 3 commits ago we restrict shortcodes to a subset of
the POSIX Portable Filename Character Set, therefore
this can never have a directory component.
E.g. *key’s emoji URLs typically don’t have file extensions, but
until now we just slapped ".png" at its end hoping for the best.
Furthermore, this gives us a chance to actually reject non-images,
which before was not feasible exatly due to those extension-less URLs
As suggested in b387f4a1c1, only steal
emoji with alphanumerc, dash, or underscore characters.
Also consolidate all validation logic into a single function.
===
Taken from akkoma#703 with cosmetic tweaks
This matches our existing validation logic from Pleroma.Emoji,
and apart from excluding the dot also POSIX’s Portable Filename
Character Set making it always safe for use in filenames.
Mastodon is even stricter also disallowing U+002D HYPEN-MINUS
and requiring at least two characters.
Given both we and Mastodon reject shortcodes excluded
by this anyway, this doesn’t seem like a loss.
Even more than with user uploads, a same-domain proxy setup bears
significant security risks due to serving untrusted content under
the main domain space.
A risky setup like that should never be the default.
Just as with uploads and emoji before, this can otherwise be used
to place counterfeit AP objects or other malicious payloads.
In this case, even if we never assign a priviliged type to content,
the remote server can and until now we just mimcked whatever it told us.
Preview URLs already handle only specific, safe content types
and redirect to the external host for all else; thus no additional
sanitisiation is needed for them.
Non-previews are all delegated to the modified ReverseProxy module.
It already has consolidated logic for building response headers
making it easy to slip in sanitisation.
Although proxy urls are prefixed by a MAC built from a server secret,
attackers can still achieve a perfect id match when they are able to
change the contents of the pointed to URL. After sending an posts
containing an attachment at a controlled destination, the proxy URL can
be read back and inserted into the payload. After injection of
counterfeits in the target server the content can again be changed
to something innocuous lessening chance of detection.
Else malicious emoji packs or our EmojiStealer MRF can
put payloads into the same domain as the instance itself.
Sanitising the content type should prevent proper clients
from acting on any potential payload.
Note, this does not affect the default emoji shipped with Akkoma
as they are handled by another plug. However, those are fully trusted
and thus not in needed of sanitisation.
This actually was already intended before to eradict all future
path-traversal-style exploits and to fix issues with some
characters like akkoma#610 in 0b2ec0ccee. However, Dedupe and
AnonymizeFilename got mixed up. The latter only anonymises the name
in Content-Disposition headers GET parameters (with link_name),
_not_ the upload path.
Even without Dedupe, the upload path is prefixed by an UUID,
so it _should_ already be hard to guess for attackers. But now
we actually can be sure no path shenanigangs occur, uploads
reliably work and save some disk space.
While this makes the final path predictable, this prediction is
not exploitable. Insertion of a back-reference to the upload
itself requires pulling off a successfull preimage attack against
SHA-256, which is deemed infeasible for the foreseeable futures.
Dedupe was already included in the default list in config.exs
since 28cfb2c37a, but this will get overridde by whatever the
config generated by the "pleroma.instance gen" task chose.
Upload+delete tests running in parallel using Dedupe might be flaky, but
this was already true before and needs its own commit to fix eventually.
The lack thereof enables spoofing ActivityPub objects.
A malicious user could upload fake activities as attachments
and (if having access to remote search) trick local and remote
fedi instances into fetching and processing it as a valid object.
If uploads are hosted on the same domain as the instance itself,
it is possible for anyone with upload access to impersonate(!)
other users of the same instance.
If uploads are exclusively hosted on a different domain, even the most
basic check of domain of the object id and fetch url matching should
prevent impersonation. However, it may still be possible to trick
servers into accepting bogus users on the upload (sub)domain and bogus
notes attributed to such users.
Instances which later migrated to a different domain and have a
permissive redirect rule in place can still be vulnerable.
If — like Akkoma — the fetching server is overly permissive with
redirects, impersonation still works.
This was possible because Plug.Static also uses our custom
MIME type mappings used for actually authentic AP objects.
Provided external storage providers don’t somehow return ActivityStream
Content-Types on their own, instances using those are also safe against
their users being spoofed via uploads.
Akkoma instances using the OnlyMedia upload filter
cannot be exploited as a vector in this way — IF the
fetching server validates the Content-Type of
fetched objects (Akkoma itself does this already).
However, restricting uploads to only multimedia files may be a bit too
heavy-handed. Instead this commit will restrict the returned
Content-Type headers for user uploaded files to a safe subset, falling
back to generic 'application/octet-stream' for anything else.
This will also protect against non-AP payloads as e.g. used in
past frontend code injection attacks.
It’s a slight regression in user comfort, if say PDFs are uploaded,
but this trade-off seems fairly acceptable.
(Note, just excluding our own custom types would offer no protection
against non-AP payloads and bear a (perhaps small) risk of a silent
regression should MIME ever decide to add a canonical extension for
ActivityPub objects)
Now, one might expect there to be other defence mechanisms
besides Content-Type preventing counterfeits from being accepted,
like e.g. validation of the queried URL and AP ID matching.
Inserting a self-reference into our uploads is hard, but unfortunately
*oma does not verify the id in such a way and happily accepts _anything_
from the same domain (without even considering redirects).
E.g. Sharkey (and possibly other *keys) seem to attempt to guard
against this by immediately refetching the object from its ID, but
this is easily circumvented by just uploading two payloads with the
ID of one linking to the other.
Unfortunately *oma is thus _both_ a vector for spoofing and
vulnerable to those spoof payloads, resulting in an easy way
to impersonate our users.
Similar flaws exists for emoji and media proxy.
Subsequent commits will fix this by rigorously sanitising
content types in more areas, hardening our checks, improving
the default config and discouraging insecure config options.
The default refresh interval of 1 day is woefully inadequate here;
users expect to be able to add the alias to their new account and
press the move button on their old account and have it work.
This allows callers to specify a maximum age before a refetch is
triggered. We set that to 5s for the move code, as a nice compromise
between Making Things Work and ensuring that this can't be used
to hammer a remote server
Mastodon at the very least seems to prevent the creation of emoji with
dots in their name (and refuses to accept them in federation). It feels
like being cautious in what we accept is reasonable here.
Colons are the emoji separator and so obviously should be blocked.
Perhaps instead of filtering out things like this we should just
do a regex match on `[a-zA-Z0-9_-]`? But that's plausibly a decision
for another day
Perhaps we should also have a centralised "is this a valid emoji shortcode?"
function
