When built with the ghcdebug flag and started with --debug=-1 (or -2
to pause at startup, or -3 to pause before exit), hledger can be
controlled by ghc-debug clients like ghc-debug-brick or a custom
ghc-debug query script.
Also, refactor version string code.
Changes:
1. rename the sandstorm "manage" permission to "edit"
(old permission names: view, add, manage;
new permission names: view, add, edit).
Rationale: "edit" best describes this permission's current powers, to users and to operators.
If we ever added more manager-type features we'd want that to be a new permission,
not a rename of the existing one (which would change the powers of existing users).
2. rename the sandstorm roles for consistency with permissions
(old role names: viewer, editor, manager;
new role names: viewer, adder, editor)
Rationale: it's needed to avoid confusion.
3. add a new option: --allow=view|add|edit|sandstorm (default: add).
'sandstorm' sets permissions according to the X-Sandstorm-Permissions header.
Drop the --capabilities and --capabilities-header options.
Rationale: it's simpler and more intuitive.
4. replace "capability" with "permission" in ui/docs/code.
Rationale: consistent with the above, more familiar.
This fixes a reported Stored XSS vulnerability in toBloodhoundJson by
encoding the user-controlled values in this payload into base64 and
parsing them with atob.
In my exploration of the vulnerability with various payloads I and
others crafted, it would appear that this is the only available XSS in
hledger-web in relation to stored accounts and transaction details. If
there is other parts of the UI which may contain user-controlled data,
they should be examined for similar things. In this instance,
protections provided by yesod and other libraries worked fine, but in a
bit of code that hledger-web was generating, the user could insert a
</Script> tag (which is valid HTML and equivalent to </script> but not
caught by the T.Replace that existed in toBloodhoundJson) in order to
switch out of a script context, allowing the parser to be reset, and for
arbitrary JavaScript to run.
The real fix is a bit more involved, but produces much better results:
Content-Security-Policy headers should be introduced, and using
sha256-<hash of script> or a different algorithm, they should be marked
as trusted in the header. This way, if the (in-browser) parser and
hledger-web generator disagree on the source code of the script, the
script won't run. Note that this would still be susceptible to attacks
that involve changing the script by escaping from the string inside it
or something similar to that, which can be avoided additionally by using
either the method used in this commit, or a proper JSON encoder.
The second approach has the advantage of preventing further XSS, to the
extent specified above, in practice, a combination of both should be
used, b64 for embedded data and the CSP sha256-hash script-src over
everything else, which will eliminate all injected or malformed script
blocks (via CSP), in combination with eliminating any HTML closing tags
which might occur in stored data (via b64).
This vulnerability appears to have been first introduced when
autocompletion was added in hledger-web, git tag hledger-0.24, commit
hash: ec51d28839
Test payload: </Script><svg onload=alert(1)//>
Closes#1525
We officially support GHC 8.6+ (and 8.8+ for hledger-web) now.
Hackage matrix builder shows all packages building successfully
with GHC 8.4+, somehow, so we'll adjust the base bound to
allow that but prevent any attempts to build with older GHCs,
