YARD static cache reads raw traversal paths before router sanitization

Summary

YARD's static cache lookup reads a request path before the router's path cleanup runs. When a server is configured with a document root, a traversal path such as /../yard-cache-secret.html is joined against that root and can return a readable sibling .html file outside the intended static tree.

The potential security risk seems low, as only html-ending files can be read, but still the risk of reading arbitrary html files is a confiendtiality issue in itself, which is why we decided to report. Please let us know if this is out of your project's scope.

Details

The --docroot CLI option stores the configured directory in server_options[:DocumentRoot] at lib/yard/cli/server.rb:198, and adapter initialization copies that value into adapter.document_root at lib/yard/server/adapter.rb:76. For Rack requests, RackAdapter#call builds a request object from the Rack environment at lib/yard/server/rack_adapter.rb:58 and passes it to router.call(request) at lib/yard/server/rack_adapter.rb:60. Router#call then stores the incoming request at lib/yard/server/router.rb:55 and invokes check_static_cache before normal routing at lib/yard/server/router.rb:56. Inside check_static_cache, the only initial guard is that adapter.document_root is present at lib/yard/server/static_caching.rb:35; the cache path is built from File.join(adapter.document_root, request.path.sub(/\.html$/, '') + '.html') at lib/yard/server/static_caching.rb:36, without cleaning .. components first. If that resolved path is a regular file, File.file? accepts it at lib/yard/server/static_caching.rb:38 and the file bytes are returned as a 200 HTML response at lib/yard/server/static_caching.rb:40. The later route sanitizer in final_options uses File.cleanpath(...).gsub(...) at lib/yard/server/router.rb:181 and lib/yard/server/router.rb:182, but a static-cache hit returns before that code is reached.

PoC

poc.zip

bash ./poc/run.sh

expected output:

run 1: exit=0 timed_out=False duration=0.08s matched=True phase=oracle fingerprint='YARD_STATIC_CACHE_PATH_TRAVERSAL'
run 2: exit=0 timed_out=False duration=0.08s matched=True phase=oracle fingerprint='YARD_STATIC_CACHE_PATH_TRAVERSAL'
run 3: exit=0 timed_out=False duration=0.08s matched=True phase=oracle fingerprint='YARD_STATIC_CACHE_PATH_TRAVERSAL'

The YARD_STATIC_CACHE_PATH_TRAVERSAL fingerprint is emitted only after the PoC observes a 200 static-cache response whose body contains the sibling file outside the configured document root. A setup failure, syntax failure, or cache miss would not print this oracle and would not demonstrate this traversal read.

Impact

A remote unauthenticated HTTP client who can reach a YARD documentation server with DocumentRoot/--docroot enabled can request .html paths containing parent-directory components and receive readable matching files outside the configured document root. The required guards are narrow: adapter.document_root must be set, the traversed target must exist as a regular readable file, and the target must be reachable through the implementation's forced .html suffix. Those requests bypass the later final_options path cleanup because the cache check runs first. The resulting severity class is information disclosure: response bodies can contain off-root .html file contents, but this path does not show write access, code execution, or arbitrary files without the .html constraint.