ARIA voice-native 3D SOC platform launches under source-available license

ARIA is a voice-native security operations cockpit built on a Node.js backend with a React 19 + Three.js frontend that can run as a web UI, Electron desktop app, or containerized service. The entire interface is rendered as a navigable 3D galaxy where sectors represent capability domains and nodes represent panels; a three-tier color system (blue nominal, amber threat, red critical) propagates live from scan and telemetry state so that posture is legible from across a room. Operators can navigate by clicking a node, speaking a destination, typing a command, or allowing ARIA to route automatically when a critical event fires.

The core design principle is the trust ladder: autonomy is not a global on-off switch but a per-capability ladder where each rung—approval, assisted, auto, full_auto—represents how much independent action ARIA is allowed to take. Critically, the AI model cannot promote itself; promotion requires an operator's explicit decision after observing measured outcomes, and demotion is immediate on a bad outcome. Every promotion, demotion, approval, and action is written to an append-only audit log with a retention envelope. This structure inverts the typical security-tool pattern in which the tool acts and then the operator is asked to trust it; ARIA requires the operator to grant trust in stages based on demonstrated performance.

Speech-to-text is always local via Whisper (through @xenova/transformers); voice never leaves the machine. Text-to-speech is an optional cloud call through ElevenLabs, and the developer notes that Piper/Kokoro is the intended replacement to complete local sovereignty. The platform ingests telemetry from cloud, identity, code, and network surfaces through connectors for GitHub, AWS, Okta, Snyk, Azure AD, VirusTotal, and Elastic Security. It performs AI Security Posture Management (AI-SPM) by inventorying AI assets, secrets, model infrastructure, endpoints, and identity links, then mapping attack paths and blast radius across them. Connector credentials are encrypted at rest with AES-256-GCM.

For regulated environments, ARIA offers a LOCAL mode that routes reasoning against an OpenAI-compatible endpoint (Ollama, vLLM, LM Studio, llama.cpp) running on the operator's own infrastructure. The code enforces sovereignty: a module called server/localLlm.mjs explicitly refuses to send a prompt to any host outside loopback or a private IP range, rejecting public addresses unless an environment variable ARIA_LOCAL_LLM_ALLOW_REMOTE is explicitly set to true. LOCAL mode never silently falls back to a cloud provider even when a cloud API key is sitting in the environment; if the local engine is down, the operator sees a clearly labeled degraded summary and the reason, not a quiet egress. The server binds to loopback by default and is not designed to face the internet.

The quickstart requires Node.js ≥ 22 and no API keys; ARIA starts with local-only features and every AI provider disabled but visibly labeled. Cloning the repository, running npm install, copying .env.example to .env.local, generating an encryption key with node scripts/generate-encryption-key.mjs, and running npm run start:local opens the web UI at http://127.0.0.1:5173 with development credentials (operator/aria). Alternatively, npm run aria launches the Electron desktop app, npm run server runs the API server alone, or docker compose up containerizes the entire stack. For Docker, setting ARIA_LOCAL_LLM_URL=http://host.docker.internal:11434/v1 reaches a model server on the host machine.

The developer has published an unedited engineering audit naming exactly which subsystems are production-grade and which are simulated. Network scan depth is real for ARP-table reads and BLE scanning but does not include a full active port or vulnerability scan. Keyword search is implemented but vector search (semantic memory) is not. Multi-tenancy has tenant IDs flowing through the system but isolation is not enforced on every store; the platform is single-tenant desktop-use only. The approval queue is file-backed and does not survive across multiple server instances. Audit log rotation is capped per file with a retention envelope but rotation itself is not yet implemented. The developer notes this is not a hardened commercial product and that parts are demo-grade; the full status with file paths and line numbers is in ROADMAP_AND_LIMITATIONS.md, which the developer recommends believing over the README if they disagree. The platform is published under Business Source License 1.1, making the code source-available for evaluation and research but not open source.

ARIA addresses a structural problem in AI security tooling: most platforms make autonomous decisions and then ask operators to trust them retroactively. By inverting this dynamic—requiring autonomy to be earned and immediately revocable—the platform applies human oversight to the most consequential decisions. The four-rung trust ladder is the mechanism: a capability only gains autonomy if an operator explicitly promotes it after observing good outcomes, and a single bad decision triggers immediate demotion. This design is particularly relevant for regulated industries where audit trails and demonstrable control are not optional requirements but legal mandates.

The sovereignty guarantee is enforced in code rather than promised in documentation, which matters. The LOCAL mode explicitly refuses to send prompts outside loopback or private IP ranges; there is no silent fallback to cloud providers even when API keys are configured. For environments where data residency is legally non-negotiable, this removes a class of risk that typical SaaS security tools cannot address.

The developer has been unusually transparent about the product's maturity. Rather than publishing a polished feature grid, the ROADMAP_AND_LIMITATIONS.md document names exact file paths and line numbers for each subsystem's status—production-grade or simulated. Local speech-to-text via Whisper is shipped, but text-to-speech is still a cloud call, so LOCAL mode is reasoning-sovereign but not yet speech-sovereign. Multi-tenancy isolation is not enforced, and the approval queue does not survive across multiple server instances. The platform is unsuitable for hardened production deployment today, but the engineering honesty and the source-available license reduce the risk of adopting it in lower-stakes environments.

No comments yet. Be the first to share your thoughts!