🔒 Security & Confidentiality Architecture
For enterprise business owners, chief financial officers, and agency leads, data privacy is non-negotiable. AMBA was engineered from day one on a Local-First, Privacy-First paradigm.
📊 Non-Technical Visual Data Flow: Your PC & The AI Cloud
To give businesses complete peace of mind, the diagram below visually explains where your data lives and what actually happens when AMBA is working:
- ›Excel, Word, PPTX, PDF files
- ›Financial ledgers & payroll
- ›OS Keychain credentials
- ›SOURCE FILES STAY ON YOUR DISK
- ›Reads files in-place
- ›Deterministic math tools
- ›Writes to your workspace
- ›Strictly no open ports
- ›Processes ephemeral prompt text in-memory to answer tasks
- ›Retention and training follow your provider's API terms
- ›AMBA never uses your data to train models
1. Zero Cloud File Storage
Your documents, spreadsheets, analytical SQLite tables, and vector embeddings remain on your laptop. AMBA never uploads copies of your work to its servers. AMBA's small cloud service only handles the free trial relay, feedback you choose to send, crash reports, update checks and the template catalog — see the Privacy Policy for exactly what each sends.
2. Ephemeral Frontier LLM Reasoning Only
When you run a task, AMBA extracts only the necessary contextual text snippets or SQL query results required to answer your prompt and sends them via outbound HTTPS to your chosen LLM provider (OpenAI, Anthropic, or Google).
- Your enterprise data is processed in-memory by the LLM provider under enterprise API privacy terms.
- It is never used to train public models.
3. Non-Technical Permissions & Security Matrix
Every project includes an intuitive Permissions & Security tab in Project Settings, allowing non-technical business users to configure exact operational boundaries:
- Filesystem Sandbox: Constrains agents strictly to your project folder, and Sources (
sources/) inside it are read-only: agents can never change, move or delete a file you supplied. Path traversal outside the designated folder is physically blocked by the Rust policy engine. - Web Search & Market Data: Controls outbound internet research. For air-gapped financial reconciliation (e.g. payroll audit, invoice verification), web access is disabled by default.
- Terminal Execution: Running OS shell commands is blocked by default to ensure the AI cannot execute unverified commands on your machine.
- External MCP Plugins: Explicit opt-in for Model Context Protocol integrations and third-party tools.
4. Cross-Lingual Reasoning Boundary (ADR-020)
To prevent prompt injection, parameter corruption, and math hallucination in multi-language environments:
- Internal reasoning and arithmetic steps in
<thought>blocks are conducted strictly in Technical English. - All user-facing communications and final deliverables are synthesized natively in your chosen project locale (
en,vi,ja,ko).
5. OS Keychain Only
- API keys and connector credentials are never stored in
.envfiles, git repositories, logs or plaintext files. - Secrets live only in the operating system's credential store:
- macOS: Keychain
- Windows: Credential Manager
- Linux: Secret Service over D-Bus (GNOME Keyring, KWallet 5.97 or later, or KeePassXC with Secret Service enabled)
- After saving, AMBA reads the key back and reports success only if the keychain kept it. Pasted keys are trimmed; line breaks, invisible characters and non-ASCII characters are rejected.
- There is no fallback store. If no keychain is available (for example Linux without a running keyring), saving fails with a clear message and Settings → System Diagnostics shows the keychain backend and whether it persists.
6. Deterministic Math, Reviewed by You
- LLMs are prone to arithmetic errors and invented numbers.
- AMBA's financial templates direct calculations to deterministic tools (SQL over local tables, spreadsheet formulas) instead of the model's mental arithmetic, and AMBA checks that promised deliverables were written to disk. The model still chooses what to compute, so review figures before you rely on them.