Security questionnaire
Your vendor review, filled in.
Pre-filled answers to the questions SIG Lite, CAIQ and most firms' own vendor forms ask, for the Private AI Workspace. Copy them into your form. If a question isn't here, email support@clientsafeai.com and we'll answer it in writing. Last reviewed October 5, 2026.
Where something isn't in place, the answer says Not yet or Partly rather than being left out.
A. Company and oversight
Who is responsible for information security?
The founder, who builds and runs ClientSafeAI. There is no separate security team, so the same person answers your security questions at support@clientsafeai.com.
Who at ClientSafeAI can access our server?
One person, the founder, to install updates and fix faults. No other staff member or contractor has access. We don’t open your documents unless you ask us to while helping you.
Do you have written security policies?
YesA short internal set covers access control, change management and backups, with an incident response plan (H1), reviewed every October. The commitments that bind us are public: this page, the Security page, the Privacy Policy and the Data Processing Agreement. It isn't a certified ISMS.
Do you have a SOC 2 report or an ISO 27001 certificate?
Not yetClientSafeAI itself hasn't been audited yet. The providers it runs on have been: see B4.
Has the service had an independent penetration test?
Not yetNo outside penetration test has been done yet.
B. Hosting and providers
Where is our data stored and processed?
On a dedicated server in the United States, rented from GPU Mart: the app, its database, your files, OCR, transcription and the AI model. The encrypted off-site copy of its nightly backup is stored with Backblaze in the United States, or with your firm’s own S3-compatible storage if you prefer.
Is any infrastructure shared with other customers?
NoYour server, database and AI model serve your firm only, and nothing is pooled into a shared index. The one thing shared is our own server, which runs billing. For your Workspace it holds your plan status, your disk’s encryption key (itself encrypted) and each night’s backup health report, never a document, question or answer.
Which third parties are involved?
GPU Mart rents us the server and controls physical access to it. Backblaze stores the encrypted backup copy. Cloudflare provides DNS for your server's address only: your traffic doesn't pass through it. ZeptoMail (Zoho) sends invitation and password-reset emails, and Stripe handles billing. Every provider is listed in the Data Processing Agreement.
Are those providers independently audited?
Cloudflare holds SOC 2 Type II, ISO 27001 and ISO 27701 certification. Backblaze holds a SOC 2 Type 2 report. GPU Mart states that its US data centers are SOC-certified. Each provider’s reports come from that provider, under its own terms.
Is our data sent to any AI, OCR or speech-to-text provider?
NoThe AI model, OCR and transcription all run on your server.
C. Encryption and data protection
Is data encrypted in transit?
YesEvery connection to your server uses TLS, with a certificate the server gets from Let’s Encrypt, and browsers are told to use HTTPS only (HSTS). Your traffic goes straight to your server, not through Cloudflare or through us. Calls from your server to ours also use TLS. Inside the server, its services talk over a private network that isn’t reachable from outside.
Is data encrypted at rest?
YesYour files, database and on-server backups sit on an encrypted disk (LUKS2). The off-site backup is encrypted on your server before it leaves (restic, AES-256).
Who holds the encryption keys?
The disk key is held by us, encrypted, on our own server. Your server fetches it over TLS each time it starts and never writes it to its own disk, so a disk taken out of it can’t be read. A recovery passphrase that unlocks the disk without us is given to your firm’s Owner. The backup key is held by your firm and by us, never by the storage provider.
Is our data used to train any model?
NoNothing you upload or ask is used to train, fine-tune or evaluate any model, on any plan. There is no setting for it because there is nothing to opt out of.
D. Access control
How do our people sign in?
Email and password. Passwords must be at least 10 characters and are stored only as bcrypt hashes.
Is multi-factor authentication available?
YesEach person can turn on two-factor authentication with an authenticator app. The Owner can require it for everyone. Until someone sets it up, the server refuses everything except the setup itself.
Do you support single sign-on (SAML or OIDC)?
Not yetSign-in is by email and password, with optional two-factor authentication.
Can we limit who sees what?
YesThree roles: Owner, Member and Read-only. A matter can also be walled, so its pages are never retrieved for anyone outside it. The wall is applied inside the search itself, and switching walls off is recorded in Activity.
How are sessions managed?
Signing in sets an HttpOnly cookie that lasts up to 30 days. The Team page lists active sessions, and any one of them, or everyone, can be signed out.
Is there protection against password guessing?
YesSign-in and password reset allow 10 attempts a minute per network address. After 10 wrong passwords or codes in a row, the account is locked for 15 minutes, and the lock is recorded in Activity. Two-factor authentication covers a guessed password.
How does ClientSafeAI access our server?
PartlyOver SSH with keys only: signing in with a password is turned off. There is no second factor on that access yet.
Are administrative actions logged?
YesThe Activity log records uploads, removals, deletions and every admin change: team, roles, settings and billing. Nothing in the app can edit or delete an entry. Sign-ins to the server itself are recorded in its system log.
E. Software development
What checks must a change pass before it is released?
A secrets scan, a vulnerability scan of each container image (a fixable critical or high finding blocks the release), Dockerfile checks, type checks, lint, unit and integration tests, and end-to-end tests against the real stack.
Are code changes peer-reviewed?
NoOne person writes the code, so there is no second reviewer. Every release has to pass the checks in E1 instead.
How are vulnerable dependencies handled?
PartlyThe image scan blocks any release with a fixable critical or high vulnerability. Dependencies are updated by hand when advisories are published. There is no automated update bot yet.
How are releases built and delivered?
Each release is a numbered version, built only after every check passes, and published as private container images. The app’s images are built on base images pinned to exact versions, and the other images your server runs (the database, the AI model server and the web server) are pinned by digest. Your server pulls them with its own read-only token.
How do we report a vulnerability?
Email support@clientsafeai.com. The same address is in our security.txt.
F. Server operations
Which network ports are open?
Three: SSH (keys only), port 80 so the server can get its TLS certificate, and 443 for the app. The database and the AI model aren’t reachable from outside the server.
Are security updates applied?
YesOperating-system security updates install automatically every day, without a reboot. The GPU driver and Docker are held back and updated by hand, since an unplanned change there could stop the AI model. App updates are installed as numbered releases.
Does the server run antivirus software?
NoUploaded files are only read, for their text, by OCR or by transcription. Nothing uploaded is ever run.
How is the server monitored?
PartlyYour server checks in with ours every five minutes and reports each night’s backup. Ours flags a server whose last good backup is more than 36 hours old, whose disks are degraded, or whose off-site copy failed. A flag emails us automatically within 15 minutes. We don’t yet get an automatic alert if your server stops checking in.
How long are logs kept?
Activity entries are kept for the life of your Workspace. Server and app logs stay on your server. Each service’s log is capped at 100 MB, oldest lines dropped first, so there is no fixed retention period.
G. Backups and continuity
How often is our data backed up, and for how long?
Every night, at about 2:30 server time. The database and files are kept as 14 daily and 8 weekly copies, both on the server and off it.
Are restores tested?
YesRestoring from the off-site copy is a single command, and we rehearse it. The last rehearsal, in October 2026, matched every database row and file. Restoring onto replacement hardware hasn’t been rehearsed yet.
How much data could be lost, and how fast is it restored?
Up to one day of changes, since backups are nightly. We don’t commit to a recovery time yet.
What if a disk fails?
The server keeps its data on two mirrored disks (RAID1), so one failed disk loses nothing.
What if ClientSafeAI’s own server is down?
Your Workspace keeps working. It needs our server only when it restarts, to fetch its disk key. If ours can’t be reached then, the disk can be unlocked by hand with the recovery passphrase.
What if ClientSafeAI stops operating?
The Owner can download everything as one ZIP at any time, from Billing → Export. If your firm keeps the off-site backups in its own storage, they sit in your account, and your firm holds their key.
H. Incidents
Do you have a written incident response plan?
YesWritten in October 2026: who does what, how an incident is graded, the first hour’s steps, how each kind of threat is contained, and the notice in H2. We walk through it once a year.
Will you tell us about a breach?
YesWithout undue delay, and within 72 hours, with what happened, the data involved and what we're doing about it. This is in the Data Processing Agreement.
I. Retention, deletion and privacy
How long is data kept?
Documents stay until your firm removes them. With soft delete on, a removed document can be restored for 30 days, then it is purged. Call audio is deleted as soon as it is transcribed, and transcripts follow the retention period you set (90 days by default). Every purge is logged in Activity.
What happens to our data when we cancel?
It is kept so you can come back, until you ask us to delete it. We then delete it, including its off-site backups, within 30 days of your request.
What does ClientSafeAI’s own server learn about our Workspace?
Whether your plan is paid, and each night whether the backup succeeded, its size and the health of the disks. Never a document, question, answer, file name, matter name, member list or Activity entry.
Will you sign a Data Processing Agreement?
YesThe Data Processing Agreement applies automatically, and we send a countersigned copy on request. For transfers out of the EEA, UK or Switzerland, we sign the Standard Contractual Clauses with you.
J. The AI itself
Which models are used, and where do they run?
Llama 3.1 8B Instruct answers questions, whisper.cpp (large-v3-turbo) transcribes recordings, and Tesseract reads scans. All three are open models that run on your server. They are downloaded once, at setup, at pinned versions checked against published hashes.
Are our questions and answers stored?
Yes, on your server only, as your firm’s chat history. Activity records that a question was asked, without its text.
Does the model learn from our use?
NoIt is never retrained or fine-tuned on what your firm uploads or asks.
What does our server connect to on the internet?
Our server (plan status, disk key and the backup report), your backup storage, our email provider for invitations and password resets, Let’s Encrypt to renew its certificate, and GitHub’s container registry when an update is installed. There is no analytics or telemetry.