Home / Capability boundaries
ObliviOS Capability Boundaries
What each service may take in, give out and do, and what it is denied.
Each service has a narrow interface. Denied capabilities must later become platform permission, Binder, SELinux, storage, and test policy.
- 1
Network policy
- Allowed inputs
- Authorized typed request.
- Allowed outputs
- Policy decision and connection summary.
- Data classification
- App identity/policy/network metadata.
- Privileges
- Assigned network-policy APIs.
- Explicit denials
- Packet-content inspection, arbitrary VPN config, cross-profile reads.
- Audit events
- Policy create/change/delete and denial.
- Failure behaviour
- Keep last valid policy; fail closed only for explicit block.
- 2
Worlds/profile manager
- Allowed inputs
- User-mediated lifecycle or share request.
- Allowed outputs
- Profile state and explicit share token.
- Data classification
- Profile metadata; profile-bound data.
- Privileges
- Multi-user/profile APIs.
- Explicit denials
- Silent cross-world access, shared clipboard by default, credential reads.
- Audit events
- Create/switch/delete/share approval.
- Failure behaviour
- Abort incomplete change; preserve profiles.
- 4
Secure Clipboard
- Allowed inputs
- Explicit copy/paste and policy.
- Allowed outputs
- Profile-scoped expiring item.
- Data classification
- Clipboard content for configured retention.
- Privileges
- Clipboard-service extension.
- Explicit denials
- Cross-world read, silent history.
- Audit events
- Copy, paste, expiry, denial.
- Failure behaviour
- Clear expired or invalid item.
- 5
Notification Firewall
- Allowed inputs
- Notification metadata and user policy.
- Allowed outputs
- Allow, suppress, summarize, quarantine.
- Data classification
- Metadata and content only with user policy.
- Privileges
- Notification filter scope.
- Explicit denials
- Default content export to AI, mutation outside policy.
- Audit events
- Decision, match, summary, override.
- Failure behaviour
- Deliver when unavailable unless user selects fail closed.
- 6
Package installer
- Allowed inputs
- Signed install/update request and metadata.
- Allowed outputs
- Allow, deny, quarantine.
- Data classification
- Manifest, signature, scanner findings.
- Privileges
- Installer integration only.
- Explicit denials
- Signature bypass, integrity spoofing, silent privileged install.
- Audit events
- Scan, signature, policy, approval.
- Failure behaviour
- Reject invalid package; retain previous version.
- 7
Updater
- Allowed inputs
- Signed channel metadata and artifact.
- Allowed outputs
- Verified staged update request.
- Data classification
- Release metadata and hashes.
- Privileges
- Updater integration/storage.
- Explicit denials
- GrapheneOS update URL, unsigned metadata, downgrade bypass, signing-key access.
- Audit events
- Check, download, signature, install, rollback.
- Failure behaviour
- Reject artifact; retain verified system.
- 8
Backup
- Allowed inputs
- User-approved configuration and handles.
- Allowed outputs
- Encrypted artifact and restore manifest.
- Data classification
- Data encrypted before export.
- Privileges
- Scoped backup APIs and key references.
- Explicit denials
- Plaintext cloud upload, vault export, cross-world mixing.
- Audit events
- Create, verify, restore, delete.
- Failure behaviour
- Preserve original data; remove incomplete artifact.
- 9
Vault
- Allowed inputs
- Explicit credential/passkey operation.
- Allowed outputs
- Authenticated credential result.
- Data classification
- Secrets and cryptographic references.
- Privileges
- Hardware-backed credential APIs.
- Explicit denials
- AI secret access, raw secret logging, plaintext backup.
- Audit events
- Request, user auth, key-use failure.
- Failure behaviour
- Deny and reveal no secret material.
- 10
Passwords app
- Allowed inputs
- The user's passphrase and, if switched on, a fresh strong-biometric authentication; the fields of a login screen that the autofill framework hands over; the user's own entries and export files.
- Allowed outputs
- Pixels on the screen; a chip (a key icon and one line of text, ADR-191, proposed) in the strip of the system keyboard, drawn by the system from this content; fills into the field the user chose (a username and a password, after an authentication); the user's own encrypted export file; the single generic "login exists" answer of the domain index.
- Data classification
- Secrets (passwords, notes, history); keyed hashes of the registrable domains and packages that have an account.
- Privileges
- An app with the user's uid; Keystore keys for the hardware secret, the index and the biometric wrapper; the autofill service and credential provider service roles, once the user selects them; no platform privilege.
- Explicit denials
INTERNETand every network and radio permission (a host check fails the build), backup of any kind, any exported component that returns a secret, bulk reads of the vault by other apps, the screen-lock credential as an unlock, a log line with a secret, a screenshot or recents image (every window is secure), saved instance state, fuzzy domain matching, fetching a logo, a chip for any keyboard other than the system keyboard, a chip that shows a password, a dataset filter taken from the stored value.- Audit events
- None of its own (no account, no telemetry); failed unlocks are counted locally and only throttle.
- Failure behaviour
- Locked or lost secret: no fill and no secret shown; the right passphrase never opens a vault whose hardware secret is gone, and the app says so and offers an import instead of an empty vault.
- 11
Location Anchor
- Allowed inputs
- Owner's explicit toggle action; one fresh location fix captured at toggle-on time; the owner's theft-recovery override setting.
- Allowed outputs
- A frozen anchor fix, with bounded jitter, served through the real-provider injection point (
LocationProviderManager.setRealProvider, seedocs/research/location-framework-api-verification.md) in place of a live fix while ON. - Data classification
- The owner's own current-session location state.
- Privileges
- Provider-injection privilege equivalent to
setRealProviderfor the four standard location providers; no network, no other app's data, no Binder/file/root/shell authority beyond that injection point. - Explicit denials
- Serving the anchor fix to the emergency-location path (always real, unconditional); serving it to the theft-recovery lookup unless the owner's explicit override is enabled; any persisted, timestamped record of on/off transitions; any API surface exposing past anchor state.
- Audit events
- None, by design — see
docs/superpowers/specs/2026-08-12-location-anchor-design.mdand ADR-037: a persisted history is a liability against the exact adversary this feature defends against, not an oversight. - Failure behaviour
- Mid-session service restart fails closed to no-fix rather than passing through a real fix; reboot preserves the prior on/off state and anchor point without creating a new record.
- 12
Duress Unlock
- Allowed inputs
- Owner's explicit setup action (decoy profile creation, separate challenge PIN); the exact credential typed at the primary keyguard on every unlock attempt.
- Allowed outputs
- Either a normal unlock into the real profile, or — on a decoy-credential match — a direct foreground switch into the pre-provisioned, isolated decoy profile with the real profile explicitly force-stopped.
- Data classification
- The owner's own credential-match outcome and current foreground-profile state; no third-party data.
- Privileges
- Primary-keyguard credential-check hook (dual-credential comparison); real-profile force-stop authority equivalent to
UserController.stopUser(INTERACT_ACROSS_USERS_FULL), seedocs/research/duress-unlock-api-verification.md. - Explicit denials
- Nested decoys; any persisted record of which credential was entered, when, or how often; counting a successful decoy match as a failed real-profile attempt; auto-generated decoy content; any indicator visible while the decoy profile is foregrounded.
- Audit events
- None, by design — a decoy-unlock event must be indistinguishable, after the fact, from a normal unlock; see
docs/superpowers/specs/2026-08-12-duress-unlock-design.md. - Failure behaviour
- A credential-check failure against both real and decoy profiles counts as one normal failed attempt against the real profile only; a hook failure or crash fails closed to today's single-credential unlock path and never reveals the decoy profile; reboot never preserves which profile was last shown — every unlock is decided fresh.
Read from docs/security/capability-boundaries.md.