Skip to content

Prototype platform security

Threat model for the platform vocabulary (requires, the platform action, platform.on handlers, {platform} reads; see “Platform actions and events” in the prototype specification). Specs are untrusted input: they come from an agent, from a saved file, or from an import into ProtoViewer (ProtoViewer). Each rule below names where it is enforced. The linked issue tracks each rule marked not yet enforced.

Threats

Threat Example
Exfiltration Read the clipboard or contacts into /data, then send it with request.
Spam and abuse Post notifications or open the share sheet in a loop, or at show time.
Persistence Keep a foreground service or background work alive after the prototype is dismissed.
Confused deputy Reach CtrlProxy’s accessibility privileges, or a host app’s TCC grants through the injected iOS agent.
Forged simulation Inject platform events or call outcomes from something other than the host.

Trust tiers

Tier Source Modules
MCP-direct prototype show from the connected agent Any module the runtime advertises.
MCP-saved save, loaded from the device’s saved store Same as direct.
Imported ProtoViewer import, library: true in the strict validator Non-sensitive modules only (R1).

Imported specs already lose privileged windows, host requests and AGSL shaders (prototypeStrictPolicy.ts).

Rules

R1: imported specs use no sensitive call. sensitive in the contract’s platformModules table marks calls and events that touch personal data (clipboard, share, contacts, location, camera and microphone output). The TypeScript, Kotlin and Swift strict validators refuse sensitive calls and handlers when library: true, with the call or handler’s path; the shared fixtures are in test/fixtures/prototype-spec/library-modules/. Relaxing this is an owner decision (#15636).

R2: sensitive results cannot reach a network or storage sink. A /data path written by a sensitive call’s into, and sensitive /platform state, may not be bound into a request URL, header, query or body, a channel URL or header, or storage.keys. The TypeScript, Kotlin and Swift validators refuse these bindings when library: true, reporting the sink path, as they do for secure text fields. Pointer overlap includes parents, children and repeat {index} segments; literal value bindings stay opaque. A setData, appendData, prependData or upsertData whose value is {data} or {platform} read of tainted state taints its /data destination, so copies and copies of copies are refused too.

TypeScript, Kotlin and Swift also track copies routed through state. A setState taints its key when its value is a sensitive {platform} read, or text naming a tainted data or platform placeholder or a tainted repeat row. A {state} read of a tainted key taints a data destination. It is itself a sink in a request, a {key} text placeholder or a state: storage key. A key holding a repeat placeholder matches every key sharing its literal prefix. The contract marks location fixes and auth user IDs as sensitive state. Successful MCP-direct shows include a warning for each sink, unless the expanded sink scan exceeds 4194304 bytes. Then they include one warning that the sinks are unchecked. No host path imports specs yet.

R3: side effects need a user gesture. The TypeScript validator permits share.open and notifications.post|schedule|setBadge only in onTap, onLongPress, onDoubleTap, onSubmit or user-initiated platform event chains (notifications.opened|action|reply, map.markerTap). Follow-ups inherit the chain’s gesture status; onAppear, onEndReached, timers and show-time calls are refused at the action path. The contract limits share to 5 calls per 10000 ms and notifications to 10 per 60000 ms. Both iOS and Android runtimes enforce each module’s sliding-window rateLimit with too_many_in_flight refusals. The Kotlin validator applies the gesture rule.

R4: runtime isolation. Where a module runs decides which privileges it can reach. The iOS installer derives Info.plist usage strings from requires.permissions, strips inherited usage strings and background modes, and omits UIBackgroundModes until the shared background vocabulary has an unambiguous Apple mapping. The iOS agent capability table is tested against the contract’s per-call permissions (#15642). The Android stamper rebuilds manifest permissions from the spec’s declarations. The install path refuses calls and nodes needing undeclared permissions before stamping. The CtrlProxy capability table is tested against the contract and its manifest.

  • CtrlProxy’s overlay service registers location, sensors (steps need motion / ACTIVITY_RECOGNITION), and permission-free modules. Its SavedPrototypeActivity additionally registers biometric, secureStorage, and permissions, which can prompt for POST_NOTIFICATIONS under CtrlProxy’s identity. The required runtime permissions are declared in CtrlProxy’s manifest; USE_BIOMETRIC comes from the renderer library’s manifest. Tests enforce exact inventories and explicit permission allow-lists for both hosts, including block/node-derived permissions. The debug-only test module uses synthetic permissions without accessing protected APIs.
  • The iOS prototype agent never advertises permissioned modules. It offers only modules that need no permission and no identity of their own.
  • The injected iOS agent never calls TCC APIs, and Info.plist filtering keeps usage strings out of an app that did not declare them.
  • A per-prototype disposable app derives permission and background declarations from the spec’s requires.permissions and requires.background, plus required install-time module permissions and Android’s INTERNET, and is uninstalled on delete or expiry.
  • Capability gates run before show, so an unsupported module is refused with its capability name instead of being drawn and silently ignored. This part is enforced.

R5: simulation is host-only. Simulated events and call outcomes arrive only on the authenticated host connection, never from spec content, intents or links. Imported prototypes accept them only while host attach is enabled, which is off by default. Not yet enforced (#15638).

R6: provenance. When an imported prototype uses a sensitive call or handler, ProtoViewer’s app-reserved provenance region names its modules. The region sits outside the spec’s drawable bounds, so a spec cannot cover it. Android (PrototypeLibraryProvenance in :protocol) and iOS (PrototypeLibraryProvenance in the agent core) derive the module list with the R1 walk. Under R1 a validated import lists none, so the region stays hidden until R1 is relaxed.

Out of scope

Ads and analytics SDKs, real push delivery and real payments are not part of the vocabulary, so they add no threats here.