This proposal concerns a strategy for isolating the permissions given to plugins in a plugin architecture. I'm looking for insight into whether this design can succeed in isolating those permissions.
Access to the entire site is controlled via OAuth2. That is, on first logging in, a user completed the oauth flow. On successful sign-in, a httpOnly JWT ID token is set in the user's browser for the domain. This ID token is required for every request. It can also be exchanged for AWS credentials (but not in the browser, because, as httpOnly, it's not accessible to JS running on the page). The ID token can be associated with multiple roles; that is, a single token may be exchangeable for any one of several roles. Let's assume that the ID token does not itself have permissions to take any actions.
For the purpose of this discussion, a plugin consists of an arbitrary directory of files to be served to a browser over HTTPS. The plugin authors would not be allowed to set response headers for these pages--the CSP header, specifically, would restrict the plugin capabilities. To prevent cross-plugin attacks, the CSP for each plugin could be narrowed to only allow requests to whitelisted API paths, even within the host domain.
Plugin authors would specify the permissions required by their plugin. On plugin installation, a role with those permissions would be created, and it would be included in the set of roles that can be given in return for the ID token of an authorized user. The goal of this design is that the code written by the plugin author should only be able to use the permissions assigned to that plugin, not the permissions of any other roles for which the user's token may be eligible.
All the plugin directories will be hosted within a single root directory; for the purpose of this discussion it will be called /plugins/. So a blog plugin might include the pages:
/plugins/blog/index.html /plugins/blog/js/main.js /plugins/blog/css/styles.css
My proposal is to write a get-credentials endpoint that exchanges the JWT ID token for AWS credentials. This endpoint would be configured with a map of plugin names to roles. On receiving a request, the endpoint would use the Referer header to determine which page sent the request. It would look up the appropriate role based on the page prefix (i.e. a request originating from plugins/blog/index.html would map to the role for the blog plugin). The get-credentials code would then attempt to exchange the ID JWT for a set of AWS credentials with the requested role. If successful, these would be returned to the script running on the page.
I'm looking for ways that this design could fail, where failure is defined as plugin code gaining access to permissions outside those included in the requested set, such as those of a different plugin or another role for which the user's JWT is eligible. Or any way of exfiltrating the ID token (except by an agent controlling the browser; I don't think I can do anything about that).
[^1]: I'd like to start out with the assumption that a malicious plugin developer could put anything in the plugin. If there turn out to be specific things it's necessary to guard against, that will be good to know.
[^2]: Let's assume that a user is placed in groups that give access to different roles. In this system, it's likely that there will be exactly one user (the system owner) and that they will have all the available permissions.