Skip to content

Security: Aqta-ai/spectra

Security

Report a vulnerability

SECURITY.md

Security policy and distribution scope

Reporting a security issue

If you find a vulnerability in the code published in this repository (the Spectra frontend, backend, the Spectra Bridge Chrome extension, or any of the supporting scripts), please do not file a public GitHub issue. Instead, email security@aqta.ai with:

  • A description of the issue
  • Steps to reproduce
  • The version / commit you tested against
  • Your name and contact information so we can credit your disclosure if you wish

We aim to acknowledge reports within 5 working days and to provide a timeline for remediation within 14 days. We do not currently run a paid bug-bounty programme; responsible disclosures will be acknowledged in release notes.

If the issue concerns the live application at https://spectra.aqta.ai, the production WebSocket endpoint, or the Spectra Bridge browser extension as distributed by Aqta Technologies, please flag clearly so we can prioritise.

Threat model

Spectra is a voice-first browser agent. The components most relevant to security are:

  • Screen capture — frames stream from the user's browser to the Gemini Live API while a session is active. Frames are JPEG-compressed and transmitted over WSS/TLS. They are not stored on any server, but they may contain sensitive content (passwords, private documents, personal data) depending on what the user has on screen.
  • Voice audio — PCM 16 kHz audio streams to the Gemini Live API while the user is speaking. Same transit and storage guarantees as screen frames.
  • Browser control — the Spectra Bridge Chrome extension can click, type, scroll, and navigate on behalf of the user. It connects only to the Spectra backend and only when the user starts a session.
  • API credentials — GOOGLE_API_KEY (or Vertex AI GOOGLE_CLOUD_PROJECT credentials) are required by the backend to reach the Gemini Live API. These are server-side only and never exposed to the frontend or extension.

Reasonable threats to report include: WebSocket session hijacking, prompt-injection escapes that cause the agent to act outside the user's intent, credential leakage through error messages or logs, extension permission escalation, and cross-origin request abuse.

What's in scope of this repository

The code in this repository is the public open-source release of Spectra. It is licensed under Apache 2.0. You may inspect it, fork it, cite it, and reuse it under the terms of the licence. Specifically the following are open:

  • frontend/ — Next.js 14 application (landing page, guide, privacy, overlay accessibility analyser)
  • backend/app/ — FastAPI WebSocket server, Gemini Live session orchestration, agent toolset (click, type, scroll, navigate, describe, snapshot, diff)
  • backend/app/agents/ — system instructions, prompt templates, narration filtering, action validation
  • backend/app/overlay/ — page analysis backend that powers /overlay
  • backend/tests/ — unit and integration tests covering core logic
  • spectra-bridge/ — the Chrome extension source (manifest, background service worker, content scripts)
  • Dockerfile, run.sh, deploy-production.sh — local-development and deployment scripts
  • ARCHITECTURE.md, OFFLINE_MODE.md, CONTRIBUTING.md, PRIVACY.md — documentation

What's NOT in scope of this repository

The following components of the Spectra platform are not distributed here and are not part of the open release:

Component Why it's closed Path to access
Production API keys and credentials (GOOGLE_API_KEY, Vertex AI service-account JSON) Operational secrets; supply your own when self-hosting Provision your own via Google AI Studio or Google Cloud
Production Cloud Run configuration (custom domain, IAP, IAM bindings, secret-manager wiring) Operational security boundary Closed
Internal monitoring and analytics Operational telemetry Closed
Hosted-version infrastructure as code (Terraform, deployment scripts specific to Aqta's GCP project) Operational security boundary Closed
Spectra Bridge published listing on the Chrome Web Store (when published) Distribution requires Aqta's developer account Use the source in this repo via Load unpacked, or wait for the Web Store listing
Internal pitch material, grant applications, business strategy documents Commercial-confidential Closed

What the live hosted version does and doesn't do

spectra.aqta.ai runs the same code published here, configured against Aqta's GOOGLE_API_KEY. The hosted version:

  • Streams screen frames and audio to the Gemini Live API only while a session is active
  • Does not store frames, audio, transcripts, or chat history server-side
  • Does not run any analytics, tracking pixels, or third-party data collection
  • Does not require an account or sign-up
  • Does not have the local Gemma 4 / Ollama offline pipeline available; that requires running Spectra on your own machine

Full details in PRIVACY.md.

If you want to use Spectra in research, accessibility audits, or partnerships

  • For self-hosted use (institutions that need Spectra on their own infrastructure for compliance): clone this repo, follow the Quick Start, supply your own Google API key. No partnership needed.
  • For accessibility audit work at scale (e.g. EU Accessibility Act compliance audits): the Overlay tool at /overlay is a useful starting point. For higher-volume or institutional engagements, contact partnerships@aqta.ai.
  • For research collaborations on voice-first accessibility, agentic UX, or assistive technology: contact hello@aqta.ai.

Citation

@misc{spectra_v0_3_0,
  title  = {Spectra — voice-first AI browser agent for accessibility},
  author = {Chueayen, Anya},
  year   = {2026},
  url    = {https://github.com/Aqta-ai/spectra},
  note   = {Apache 2.0 licensed.}
}

Contact

There aren't any published security advisories