This is a universal cross-site-scripting (UXSS) / cross-origin-isolation logic bug in WebKit’s Web Archive loading path. A .webarchive (application/x-webarchive, application/x-mimearchive, multipart/related) is a self-contained snapshot whose main resource can declare an arbitrary origin/URL and can contain script; the security invariant is that loading such an archive must not let its script run against, or read the persistent state (cookies, storage, credentials, permissions) of, real origins in the user’s normal (persistent) data store. The large diff is spread across many files, but the security-relevant core is isolable: it moves web-archive loads onto an isolated, non-persistent (ephemeral) WebsiteDataStore and hardens the gate that decides when an archive may load at all. In WebPageProxy::receivedNavigationActionPolicyDecision(), when the new loadWebArchiveWithEphemeralStorageEnabled preference is on and the navigation is a Use decision for a substitute web archive or a file:// URL ending in .webarchive, the code now swaps in a freshly created non-persistent data store (WebsiteDataStore::createNonPersistent()), records the replaced persistent store, forces a process swap, and marks the load as LoadedWebArchive::Yes; navigating away restores the original store. That LoadedWebArchive flag is threaded through backForwardAddItemShared(), ProvisionalPageProxy, WebBackForwardListItem (which now carries a dataStoreForWebArchive), and NetworkProcess::AddAllowedFirstPartyForCookies so cookie/first-party bookkeeping matches the ephemeral store rather than the user’s real store. Additionally, once a web archive has been loaded, WebPageProxy::queryPermission() now returns PermissionState::Denied outright, preventing archive content from gaining camera/mic/geolocation etc. DocumentLoader::disallowWebArchive() is refactored to use the new shared MIMETypeRegistry::isWebArchiveMIMEType() helper (whose isValidValue guard also avoids feeding an over-long/invalid string to the hash set) and to key the local-webarchive allowance on the AlwaysAllowLocalWebarchive setting rather than a per-loader flag. Separately, isRequestFromClientOrUserInput is now propagated through LoadParameters and only set on the WebProcess FrameLoadRequest when the UI process says the load actually originated from client/user input, so a web-content-initiated navigation cannot masquerade as user-initiated. The pre-patch invariant violation: archive content loaded into (or sharing) the persistent data store and inheriting normal permission/first-party-cookie treatment, allowing it to act with the authority of real origins — i.e. UXSS. Note: because the enumerated source files (LoadParameters.serialization.in, UnifiedWebPreferences.yaml, the API tests) are omitted, the exact new-preference default and IPC serialization are inferred from the code that references them, not shown by the diff.

This is a same-origin-policy / cross-origin isolation failure (UXSS), not a memory-corruption primitive: exploited pre-patch, malicious web-archive script runs with the authority of spoofed real origins, letting it read cookies/storage and make authenticated cross-site requests, and (before the queryPermission fix) request powerful device permissions as those origins. It is a confidentiality/integrity break confined to the WebContent process and the browsing profile rather than a path to native code execution, so no heap-grooming or ACE chain is implied by the diff. Because a web archive is a delivered file rather than a drive-by web page, exploitation typically requires getting the victim to open a crafted archive, which is why the severity is rated medium.