ORIX Astra — Security and privacy
This document states what the app does, what it defends against, and how each defence is tested. Items that could not be verified on a physical device are marked as such.
1. Data inventory
| Data | Where it lives | Leaves the device? |
|---|---|---|
| Conversations, messages, branches, prompts, presets, agent profiles and runs, usage ledger, diagnostics, memories | SQLite database in the app's Application Support directory, file protection completeUntilFirstUserAuthentication |
Only through an export the person starts |
| Attachments (images, PDFs, text) | Files named by SHA-256 beside the database | Sent to a cloud provider only when attached to a message for a cloud model, after a one-time confirmation per provider |
| Imported documents and their full-text index | Same database | Retrieved passages are sent to a cloud provider only if the agent's or chat's model is a cloud model |
| Provider API keys, custom-endpoint keys | iOS Keychain, kSecAttrAccessibleWhenUnlockedThisDeviceOnly, not synchronisable |
Sent only to their own provider, in a header |
| Model weights | Application Support/Models, excluded from backup | Never |
| Analytics, telemetry, crash reports | None collected | — |
There is no ORIX account, no ORIX backend in the request path, and no third-party SDK. The app's privacy manifest declares no tracking and no collected data types.
2. Where requests go
| Feature | Destination | When |
|---|---|---|
| Cloud chat, arena, agents on a cloud model | The selected provider (api.openai.com, api.anthropic.com, generativelanguage.googleapis.com, api.x.ai, openrouter.ai) or the person's own endpoint |
When the person sends a message to that model |
| Model lists and key checks | Same hosts | When a key is saved or the list is refreshed |
| Model catalogue and weights | www.orixlab.dev, with huggingface.co as fallback and for licences that are not mirrored |
When the person opens Models or starts a download |
Agent web_fetch |
The host in the approved request | Only after a per-host approval |
| On-device chat (MLX, Apple Intelligence) | Nowhere | — |
OpenRouter forwards requests to an upstream provider of its choosing; the app shows which one served each answer and offers the "providers that do not retain prompts" routing option.
3. Threat model and mitigations
| # | Threat | Mitigation | Test |
|---|---|---|---|
| T1 | A key leaks into a log, an error message, a diagnostic, an export or a screenshot | Keys are read from the Keychain at request time by a closure and never stored in observable state; the key field is cleared the moment it is submitted and only a masked form (prefix••••last4) is ever shown; Redactor scrubs known key shapes, bearer tokens, credential headers and key=/token= parameters from every provider message before it becomes an AstraError; diagnostics record status, request id and timings only |
errorMessagesNeverCarryASecret, redactorRemovesKnownKeyShapes (7 key shapes × 7 contexts), provider key-leak test across errors, URLs, bodies and diagnostics for all six adapters, export test asserting no key field |
| T2 | A key is sent to the wrong host | One adapter per provider with a fixed base URL; Gemini's key goes in the x-goog-api-key header, never in the URL; custom endpoints use their own Keychain account |
Request-encoding tests per adapter |
| T3 | A development key ships in the app | The shipped app contains no key. Automated runs may inject keys through the process environment in DEBUG builds only; they are held in an in-memory store and never reach the Keychain or disk | Secret scan (Scripts/secret-scan.sh) over the repository and the built app |
| T4 | Insecure endpoint | Plain HTTP is refused by default; when explicitly allowed for one endpoint it is still limited to loopback, link-local, private and CGNAT ranges and .local hosts, matching App Transport Security's local-networking exception; credentials in the URL are refused |
EndpointPolicyTests |
| T5 | Header injection through custom endpoint headers | Header names are validated, credential-bearing names are refused, values with control characters are refused (checked on Unicode scalars: \r\n is one Character in Swift, a bug the tests caught) |
profileRefusesCredentialHeaders |
| T6 | Malformed or hostile provider stream | Incremental SSE parser with an 8 MB line cap; StreamValidator rejects out-of-order events; invalid tool-argument JSON becomes an empty object with the raw text preserved; exactly one terminal event is guaranteed |
ValidatorTests, OrchestratorTests, per-adapter mid-stream error fixtures |
| T7 | Model output that tries to run code or load remote content | Model output is rendered as Markdown by a native parser; raw HTML is shown as literal text and never interpreted; remote images are never fetched (the link is offered instead); only http, https and mailto links can be opened, anything else is inert |
Markdown model tests |
| T8 | Prompt injection through documents, web pages or tool results | Tool results reach the model inside a delimited envelope with a fixed "this is data, not instructions" preamble, with marker spoofing stripped; permissions, allowlists and budgets are enforced in code so no model output can widen them; a heuristic scanner flags suspicious content and the flag is shown in the timeline and the approval sheet | Agent injection scenarios |
| T9 | Tool misuse by an agent | Tools outside the profile's allowlist never execute; arguments are validated against the tool's JSON Schema before execution; writes, external requests and destructive actions always need an explicit approval; a denied call returns an error to the model and is recorded; steps, tool calls, tokens, cost and wall-clock are bounded; repeated and alternating calls are detected and stop the run | Registry, approvals matrix, budget and cycle tests |
| T10 | Data exfiltration through web_fetch |
HTTPS only, public hosts only (loopback, private, link-local and CGNAT refused), every redirect re-validated, URLs whose query embeds chunks of conversation text are flagged and always prompt even on a pre-approved host | SSRF and exfiltration tests |
| T11 | Path traversal and oversized or hostile files | Attachment bytes are stored under their SHA-256, never under a supplied name; imported files are capped at 20 MB, images are decoded at target size through ImageIO and re-encoded (which also drops location metadata), extracted text is capped; model downloads restrict paths to the model directory, refuse .. and absolute paths, and refuse bodies larger than the declared size |
Persistence traversal tests, download coordinator tests |
| T12 | Tampered or corrupted model weights | Every file is verified by SHA-256 against the catalogue; the bundled catalogue pins the hashes of the models the app ships with, and an online card for a known model is rejected if its hashes differ | Catalogue and download tests |
| T13 | Decompression bombs | The app never decompresses archives: models are downloaded as plain files, attachments are not unpacked | By construction |
| T14 | Low disk, low memory, thermal throttling, interrupted downloads | Disk preflight with a safety margin; conservative memory verdicts; one loaded model; unload on memory pressure; refusal to start a generation at critical thermal state; resumable range downloads that survive relaunch | Governor and download tests (injected readings); device behaviour unverified |
| T15 | Someone picks up the unlocked phone | Optional app lock with device-owner authentication (Face ID / Touch ID with passcode fallback); the lock view covers content when the app leaves the foreground; if the device has no passcode the lock is turned off and the person is told why | Unverified on device |
| T16 | Quota exhaustion and runaway cost | Arena shows an upper-bound estimate before a multi-model run and supports a budget; agents have token and cost limits; costs are labelled as estimates with the date of the prices | Cost tests |
4. Reasoning traces
The app shows only what a provider legitimately returns as a reasoning summary, or a local model's own visible thinking block. It never attempts to reconstruct hidden reasoning and never displays opaque reasoning state.
5. Clipboard
The app writes to the clipboard only on an explicit Copy action and never writes a key to it. Pasting a key reads the clipboard once, on the person's tap.
6. What agents cannot do
No shell, no process spawning, no arbitrary code execution, no access to other apps' data, no device-wide automation, no file access outside the items the person grants for a run, no file write except through the system exporter after approval. calculator and data_transform are the supported limited evaluators.
7. Transport security
All provider traffic is HTTPS with the system's certificate validation; the app does not disable or weaken it and uses ephemeral sessions without cookies or URL caches. NSAllowsLocalNetworking is the only App Transport Security exception, for user-configured servers on the local network.
8. Export compliance
The app uses only HTTPS and Apple's system cryptography (Keychain, CryptoKit SHA-256): ITSAppUsesNonExemptEncryption is false.
9. Reporting
Security reports: admin@spboucher.ai.