ObliviOS Developer portal Project site

Home / Decisions / ADR-167

Harden the identity surface of apps as designed in 2026-10-14-identity-surface-hardening-design.md.

Accepted, recorded 14 October 2026.

Decision
ADR-167
Date
2026-10-14
Status
AcceptedAccepted (owner, 2026-10-14, all recommendations of the design).
Part
Compartments
Mentioned by
ADR-173, ADR-174, ADR-206, ADR-208

1Decision

Harden the identity surface of apps as designed in 2026-10-14-identity-surface-hardening-design.md. (1) Channel A: the ro.boot.* properties that identify a unit or a production batch stop being readable by appdomain; no allow rule is added for apps, and the list is fixed only after the probe has run on stallion.

(2) Channel C: for a non-system uid the bias estimates of the uncalibrated accelerometer, gyroscope and magnetometer are folded into the readings and delivered as zero; direct sensor channels for these three types are refused to non-system uids if the plan's probe shows that they carry the bias. (3) Channel B (boot id, boot time) stays a documented residual.

(4) The model of section 11 is adopted: a per-app identity mode (Hardened by default, or Real) and, per permission class an app demands, Real, Decoy or Deny, with one policy store in system_server; channel A stays global because SELinux domains are fixed when the image is built. (5) Order: channels A and C global, then the policy store and its settings screen, then the sensor decoy, then the location decoy; each is its own plan.

(6) A separate network exit per app is not decided and not designed (D3 of the design).

2Why and evidence

A probe in two apps with different signing keys on lynx (userdebug) found that an app without a dangerous permission reads, equally in both apps, unit properties of ro.boot.* that private/domain.te:519 allows every domain, the bias estimates of three uncalibrated sensors, and the boot identity; ANDROID_ID and the DRM device identifier already differ per app; uncalibrated - bias equals the calibrated stream within the noise of two 5-second windows (lane-feasibility.md).

3Consequences

The claim is limited to one device, one userdebug image and targetSdk 36: equality across apps is measured, uniqueness across devices is not, persistence of the bias over a reboot is not, and the residual error of calibrated sensor streams is not changed. Two patches in the package privacy/identity-surface of sovereignos-platform (system-sepolicy, frameworks-native), authored by the worktree rule of AGENTS.md.

capability-boundaries.md and threat-model.md get a row each in the same change. The probe settled consequence (2) on 2026-10-14: on the baseline image an ordinary app configures a direct sensor channel (type HardwareBuffer) for all three types, so the refusal is part of the change; that the bias is written into such a buffer is not read (Java cannot read it) and stays an assumption.

A fresh review of the patches on 2026-10-14, checked in source, narrowed two points: the read rule also excludes app_zygote and native_app_zygote (app code runs there and they are not appdomain), and the sensor fold and the direct-channel refusal apply to app uids (uid % AID_USER_OFFSET >= AID_APP_START) instead of every uid but AID_SYSTEM, so native daemons and platform uids keep the unchanged events.

4Review when

A property on the list turns out to be read by a platform component; the probe shows the bias on a user build or on stallion after the change; a new upstream release changes sendEvents.

Read from docs/governance/decision-log.md, entry ADR-167.