ObliviOS Developer portal Project site

Home / Decisions / ADR-188

Protect what an app waits for from the masked tier, at run time, and give every app a way out.

Proposed, recorded 22 October 2026.

Decision
ADR-188
Date
2026-10-22
Status
ProposedProposed (owner approved the guard and its notification, 2026-10-22)
Part
Compartments, Network exit
Mentions
ADR-158, ADR-186

1Decision

Protect what an app waits for from the masked tier, at run time, and give every app a way out. (1) A third stand per app, "Direct, never hidden": the app's uids join the Exempt class, so its DNS is passed on and nothing it connects to is masked; the user sets it for a browser or a video app that the tier slows down.

(2) The speed guard in exitd: the daemon counts the bytes that a masked company sends to an app over Tor; 4 MiB or more from one company to one app within 60 s makes the daemon pass that company's names on to that app for 30 minutes (RAM only; the same effect as the user's "Show my address", applied from the next DNS answer), reports it to the exit app in the state reply, and the app notifies once per promotion and lists it in Settings with "Hide again", which releases it and keeps the company hidden for that app until the daemon restarts.

The signal is the amount of data, not time: an idle keep-alive connection of a tracker stays open for minutes, and the first connection after a cold circuit takes more than 3 s, so either would unmask nearly every company; slowness to open stays with the ladder of ADR-186, which asks the user.

(3) The deny list is no longer the only protection and grows from evidence: online-metrix.net (device fingerprinting for banks) and clickbank.net (checkout) now; the hosts that the owner's own apps contact, from the Data Map, once the tier runs on stallion. (4) The warning for Private DNS says that hosts may still see the address (the tier does not mask where DNS-over-TLS works), and the consent text names attribution and crash reporting.

2Why and evidence

Tor is 7 to 89 times slower for 10 MB and 11 to 18 times slower to start a request [measured: NET-005], so a masked host that serves content, not a beacon, is a regression the user sees; the classifier files whole domains and content delivery hosts under advertising (hwcdn.net, phncdn.com, line-scdn.net are masked rows [measured: exit-mask-list.tsv, 2026-10-22]), and a name filter finds nothing (406 of 3062 masked names contain a content-like word, 118 of them "media" of ad companies [measured: same file]).

Other breakage lists are per site and for browsers, and the lists of DuckDuckGo, Disconnect and Ghostery are non-commercial (ADR-158). The library already counts bytes per app in its copier [verified-in-source: exit-daemon/third_party/onionmasq/crates/onion-tunnel/src/proxy.rs InteractiveCopier, BandwidthCounter]; it lacks the company of the flow, which patch 0006 adds.

3Consequences

A host can unmask itself by sending 4 MiB to the app: a tracker that knows the tier could do it on purpose. The effect is bounded (30 minutes, one company, one app, RAM), visible (notification and Settings) and the same as the user's own choice; the guard is a usability protection, not an anonymity claim, and the threat model says so. The first 4 MiB of a heavy flow have gone through Tor and a long single connection is not moved.

4Review when

The device rounds show that real pages or video still lag beyond what the guard catches (then a lower threshold or a deny list from the sweep); the guard unmasks companies that the owner's traffic shows to be beacons only (then a higher threshold); a review finds a way for an app to force the promotion of a company for another app.

Read from docs/governance/decision-log.md, entry ADR-188.