Our security architecture, in plain language
- On-device analysis. All file scanning runs locally on your machine, and so does the optional AI analysis, on machines where you already run Ollama yourself. Your files and their contents are never uploaded.
- Recoverable by design. Deletions are quarantine-first or go to your OS Trash. A scan never irreversibly removes anything.
- Sensitive files are shielded. On-device checks flag documents holding personally identifiable information (PII), health records, or financial account numbers. Flagged files are excluded from bulk deletion. The claim and its test.
- Signed & verified builds. macOS builds are Developer ID–signed and notarized by Apple. Windows builds are Authenticode-signed as CN=Vaulith LLC. If your OS warns about an unsigned copy, it did not come from us.
- Minimal accounts. You can use Vaulith without an account. Sign-in exists only for purchases and licenses. It uses PKCE OAuth, and sessions live in your OS keychain.
- Payments we never see. Checkout is handled entirely by Stripe. Card numbers never touch our systems.
- Verifiable, not just promised. Watch the app’s network activity yourself while it scans — a 60-second copy-paste test with its expected output. Every falsifiable claim is catalogued with its disproving test on the claims ledger.
For reviewers and IT teams, the Security Whitepaper maps this architecture to controls: the no-egress design, the signing chain, sensitive-data detection, recoverable deletion, subprocessors, and the honest roadmap. Deploying a fleet? See Vaulith for Teams.
The Wireshark Invitation
Every cleaner says it’s private. We’d rather you check than take our word. This is a standing, public invitation to point the harshest instrument you own at our central privacy claim — and a promise about what happens if you catch us.
The claim under test
No Vaulith build from v1.0.54 (August 14, 2026) onward transmits your file names, file paths, thumbnails, or file contents off your machine — in any sign-in state: signed out, signed in, free or paid. Crash reports are path-scrubbed; if one ever carries a file name or path, that counts too. Builds before v1.0.54 carried a weaker crash-report scrubber — we hardened it, and we say so here: verify against v1.0.54 and later.
What counts as catching us
- A reproducible packet capture showing outbound traffic from the Vaulith app process transmitting any of the above, together with the steps to reproduce it.
- Against an unmodified, shipped build — a tagged release (v1.0.54 or later) from the public releases repo, verifiable against its published
SHA256SUMS.txt.
- Judged against the documented expected-traffic list: anything in the categories above counts, even if encrypted. We reproduce your steps against the same tagged release to confirm.
What we promise if you do
- We treat it as a severity-one defect: we confirm or ask questions within 48 hours, fix it, and ship the fix.
- We disclose it publicly on the claims ledger — the claim’s status changes on the page built to record exactly this.
- You get named credit (with your permission) in the disclosure and the release notes.
- Safe harbor applies (below) — good-faith research against this claim will never be met with legal action from us.
- Out of scope: modified or patched builds; traffic from other software on the machine (including Ollama’s own model downloads); responses injected by a proxy or man-in-the-middle you control; and content you yourself typed into support or feedback channels.
How to submit
Email support@vaulith.com with subject SECURITY REPORT.
Why we can invite this: the Vaulith Engine — file intelligence that runs where your files live — physically contains no network client. The claim is not a policy we enforce. It is an architecture we would have to rebuild to violate. The cheapest way to check us is the 60-second test. The invitation exists for whoever brings the heavier instruments.
Release integrity: signing, notarization, checksums, and the pinned update key
macOS signing
Developer ID–signed and notarized by Apple. Team ID: F82V4P432D. Gatekeeper verifies this before first launch — a legitimate copy needs no warnings overridden.
Windows signing
Authenticode-signed, certificate subject CN=Vaulith LLC (verified on the shipped installer). Check it yourself: right-click the installer → Properties → Digital Signatures. SmartScreen may still show a one-time new-publisher prompt while reputation accrues. The signature is the fact to trust. An unsigned “Vaulith” installer is not ours.
Checksums
Every release ships a
SHA256SUMS.txt beside its installers on
GitHub Releases. The verification one-liner and its expected output are on
verify-it-yourself.
Auto-update key
In-app updates are signed with a minisign key. Its public half is pinned inside every shipped build, and the updater refuses any update not signed by it. The key is never rotated. So even if our distribution channel were compromised, an update signed by anyone but us cannot install. The pinned public key:
untrusted comment: minisign public key: 48229AC92E104E51
RWRRThAuyZoiSJiWmG9gGgWaFLSLK6dgKnJXJXOgNIm61plLE6eJgCE0
Reporting a vulnerability
If you believe you've found a security issue in the Vaulith desktop app, our website, dashboard, or infrastructure, we want to hear about it — before anyone else does.
- Email support@vaulith.com with the subject line
SECURITY REPORT.
- Include steps to reproduce, the affected component and version, and impact as you understand it.
- We acknowledge reports within 48 hours and aim to triage within 5 business days. We'll keep you informed through to resolution.
- Machine-readable contact details:
/.well-known/security.txt (RFC 9116).
Safe harbor for good-faith research
We will not pursue legal action against researchers who: act in good faith, avoid privacy violations and service disruption, do not access or modify other users' data, give us reasonable time to remediate before public disclosure, and do not exploit findings beyond what is necessary to demonstrate them. This safe harbor explicitly covers research conducted for the Wireshark Invitation above.
Out of scope: denial-of-service testing, social engineering of our staff or users, spam, and findings that require physical access to a victim's unlocked device.
We don’t run a paid bounty program — but we credit researchers (with permission) in release notes, and we take every report seriously. This policy was last updated August 12, 2026.