Default capabilities are initially set up when a follow request is made. The Accept activity sent back from a follow request can be received by any instance. A capabilities accept activity is attached to the follow accept.
The default capabilities could be *any preferred policy* of the instance. They could be no capabilities at all, read only or full access to everything.
When posts are subsequently sent from the following instance (server-to-server) they should have the corresponding capability id string attached within the Create wrapper. To handle the *shared inbox* scenario this should be a list rather than a single string. In the above example that would be *['http://bobdomain.net/caps/alice@alicedomain.net#rOYtHApyr4ZWDUgEE1KqjhTe0kI3T2wJ']*. It should contain a random token which is hard to guess by brute force methods.
Subsequently **Bob** could change the stored capabilities for **Alice** in their database, giving the new object a different id. This could be sent back to **Alice** as an **Update** activity with attached capability.
Bob can send this to Alice, altering *capability* to now include *inbox:noreply*. Notice that the random token at the end of the *id* has changed, so that Alice can't continue to use the old capabilities.
If she sets her system to somehow ignore the update then if capabilities are strictly enforced she will no longer be able to send messages to Bob's inbox.
Object capabilities can be strictly enforced by adding the **--ocap** option when running the server. The only activities which it is not enforced upon are **Follow** and **Accept**. Anyone can create a follow request or accept updated capabilities.
## Object capabilities in the shared inbox scenario
Shared inboxes are obviously essential for any kind of scalability, otherwise there would be vast amounts of duplicated messages being dumped onto the intertubes like a big truck.
With the shared inbox instead of sending from Alice to 500 of her fans on a different instance - repeatedly sending the same message to individual inboxes - a single message is sent to its shared inbox (which has its own special account called 'inbox') and it then decides how to distribute that. If a list of capability ids is attached to the message which gets sent to the shared inbox then the receiving server can use that.
If **Eve** subsequently learns what the capabilities id is for **Alice** by somehow intercepting the traffic (eg. suppose she works for *Eveflare*) then she can't gain the capabilities of Alice due to the *scope* parameter against which the actors of incoming posts are checked.
**Eve** could create a post pretending to be from Alice's domain, but the http signature check would fail due to her not having Alice's keys.
The only scenarios in which Eve might triumph would be if she could also do DNS highjacking and:
* Bob isn't storing Alice's public key and looks it up repeatedly
* Alice and Bob's instances are foolishly configured to perform *blind key rotation* such that her being in the middle is indistinguishable from expected key changes
Even if Eve has an account on Alice's instance this won't help her very much unless she can get write access to the database.