Trust & privacy
Your data stays on your Mac.
No telemetry. Your prompts, chats, documents, audio, images, video and model files never leave your Mac unless you deliberately use a network feature. Bedrova is closed-source, so instead of asking you to trust us, we publish exactly what the app does on the network and show you how to check for yourself.
Every network call Bedrova can make
| Call | When | Where | What's sent |
|---|---|---|---|
| Model search | Only when you search or open a model in Discover | Hugging Face API | Your search terms and the model you asked about |
| Model downloads | Only when you download a model | Hugging Face CDN | The files you requested, plus your IP, as with any download |
| App update check | Manual, or if you turn on auto-update (off by default) | Bedrova update feed (Sparkle) | App version only. No usage data, no system profile |
| Runtime pack updates | Only when you check for or apply an engine update | Bedrova pack repository | Signed manifests and the pack files themselves |
| Anonymous analytics | Only if you turn it on (off by default) | Bedrova | Anonymous activation events. Never prompts, chats, files, model names or paths |
| Links you click | Only when you click an external link | That site | A standard browser request |
By default there is no background telemetry, crash-reporting, or "phone home." Inference is 100% on-device.
Threat model
What Bedrova's design protects — and what's still up to you.
Bedrova protects against
- Data exfiltration — no prompts/files/models sent anywhere by default
- Malicious model code — safetensors only, no remote code execution
- Silent updates — signed builds + signed runtime packs with rollback
- Surprise network calls — every call documented and verifiable
- Anything at all leaving — Offline Mode blocks outbound traffic outright while local work carries on
Out of scope / your responsibility
- Model content & accuracy — open models can be wrong or biased
- What you choose to download — trust the source of gated/community models
- Exposing the server — LAN/Tailscale is off by default and key-gated
- Physical device security — your Mac's disk and login
How Bedrova compares on privacy
| Capability | Bedrova (local) | Typical cloud AI |
|---|---|---|
| Prompts & files leave your device | Never | Always |
| Your data used for training | Never | Often (check ToS) |
| Account required | No | Yes |
| Works fully offline | Yes | No |
| Telemetry by default | None | Usually |
General comparison vs typical hosted AI services; always check a specific provider's terms.
Verify it yourself
- Open Little Snitch (or run
sudo tcpdump -n 'tcp port 443') and watch Bedrova while you chat, make an image, or transcribe audio. - You'll see no outbound connections during local inference — only the calls in the table above, and only when you trigger them.
Model trust
Downloaded models load as safetensors with no remote code execution (no trust_remote_code). Tokenizer, config and prompt templates are validated; new downloads are quarantined until manifest/hash/license checks pass, and trust state is shown before load.
Provenance & SBOM
Each build carries its git SHA, runtime/provider-pack manifests (signed, versioned), and component hashes, exportable in a redacted diagnostics bundle. App updates are Ed25519-signed via Sparkle; runtime packs use TUF-style signed manifests with rollback protection. Open-source components ship with their MIT/Apache attributions in-app.
Report a security issue
Email security@bedrova.com — see security.txt. We disclose and fix responsibly.
FAQ
Is Bedrova open source?
No — Bedrova is closed-source and free. Instead of asking you to trust us, we publish exactly what the app does on the network and show you how to verify it.
What leaves my Mac?
Nothing unless you act. Model search and downloads when you go looking for a model, an update check if you enable it, and anonymous analytics if you turn them on. All inference is on-device. Offline Mode blocks even those, while local work keeps running.
Can a downloaded model run code on my Mac?
No. Models load as safetensors with trust_remote_code disabled, and tokenizer/config/templates are validated. Downloaded models are quarantined until checks pass.
How do I report a vulnerability?
Email security@bedrova.com (see /.well-known/security.txt). We disclose and fix responsibly.
About this website: no analytics, no cookies, no third-party trackers, and fonts served from our own domain rather than a CDN.