Your customers’ words, kept apart and on the record.
Written for whoever reviews vendors for you. Each control says what it does, and where it is only partly in place, says so.
- Step 1: Your helpdeskRead with the access you grant: your AI’s conversations, or all of them.
- Step 2: Redacted on arrivalNames, emails, phone numbers, card numbers, keys and tokens removed before the first write.
- Step 3: Your own databaseOne per workspace, in the US or the EU, chosen at setup.
- Step 4: Models, through one gatewayThey see redacted text only, keep none of it and do not train on it.
That is the way in. Nothing reaches a customer without a person approving the exact words.
Isolation and encryption
No other customer is in your database, so there is no filter to get wrong.
- One Neon project per workspace
- Separate storage, compute, credentials and restore history. No shared table, no tenant column, no row policy that has to be right.
- Region chosen once
- United States or Europe, before the workspace exists; it cannot be moved. Model calls are processed in the US, and the agreement relies on the Standard Contractual Clauses for that.
- Zero data retention at the models
- Every model call asks the gateway for zero data retention, so it is served only by Anthropic or OpenAI under an agreement to keep no prompt or answer once it is returned (OpenAI keeps a narrow exception for safety review), and the gateway keeps none either. Neither trains on it. What we record is the usage: tokens and cost.
- Encrypted in transit and at rest
- HTTPS with HSTS for every page (SSL Labs: A+ and A across our two edge addresses, tested 15 October 2026). Every connection to a database is TLS with the server’s certificate verified. Neon encrypts storage at rest.
- Credentials encrypted with AWS KMS
- Helpdesk tokens, connection strings and webhook secrets are encrypted by KMS itself, each bound to its workspace and kind, so every decrypt is in CloudTrail. Production refuses to store a credential without it.
- A least-privilege role
- Database triggers refuse any change to the audit trail, even from the table’s owner, and workspaces are moving onto an app role that owns nothing and holds no right to change it.
Staff access
Our staff open your workspace only for a reason you can read.
- A stated reason, recorded first
- Opening a workspace’s data takes an elevation with a written reason. The record is written before the work, and the access is refused if it cannot be.
- Time-boxed
- Up to four hours for counts, one hour for anything customers wrote, for that one workspace and no other.
- Shown to you
- Every opening appears on Settings › Activity, our staff named as staff, with the reason they gave.
- Grants that lapse
- Staff access is a record with separated duties that expires within 90 days. Unusual staff access alerts us.
The audit trail
Who did what, when, and what they approved, in a record nobody can quietly edit.
- Append-only, enforced by the database
- Triggers refuse an update, a delete or a truncate, even from the table’s owner. Where the least-privilege role is in place, the database refuses them by privilege before a trigger is reached.
- Hash-chained
- Each row hashes its own fields and the row before it. Verification recomputes every row, so an edit shows at the row it was made in; concurrent writes are serialized, so the chain never forks.
- A change and its record commit together
- An approval whose record fails to write leaves no approval behind.
- Yours to export
- The audit trail and the access log download from Settings › Data. They hold who, what, when and counts — never customer text or a token.
Controls on the AI
Everything a model reads or writes is treated as untrusted. Code decides; a person approves.
- Customer text is fenced
- In every prompt, what a customer wrote is marked as data, one tagged turn per message with the tags escaped out of it. This narrows prompt injection; it does not end it, which is why the approval below is the control.
- Tool descriptions never reach a model
- A connected tool server’s descriptions are hashed and shown to a person, and read by nothing else. If one changes, the tool is suspended until a person reads it again.
- A model never chooses who or what
- Recipients, targets, the tool and its arguments are decided by code from what a person approved. Only ids that were sent to a model are accepted back.
- Citations are checked
- A verdict is kept only when the passage it quotes appears in the article it names. No citation, no verdict.
- Exact-text approval
- Nothing reaches a customer without a person approving the exact words. The approval binds an unchangeable copy of the words, and a send is held if anything it rested on moves before it goes.
Your data, your call
Everything here is self-serve on Settings › Data, and every action is in the audit trail.
- Retention you set
- As long as your plan keeps history unless you choose less (1 year on the paid plan, 90 days on Free), never under 7 days. Each night, conversations past it are deleted with everything derived from them, and deleted data leaves the database’s restore history within 6 hours.
- Export
- One person, the whole workspace, or the logs, behind a link that expires in seven days and allows three downloads.
- Delete
- Erase one person across every source, or delete one conversation. There is no bulk delete by filter, on purpose.
- Close the workspace
- Everything outward stops at once, you get a final export, and after seven days the database is destroyed and the receipt emailed to you.
- The data processing agreement
- Accepted in the product against a versioned subprocessor list; a change to either asks again.
Hardening
The defaults a reviewer checks for.
- Content-Security-Policy
- Enforced on the app and the console: scripts load only from our own origin, our sign-in provider and its bot check, and nothing may frame them; the public pages are static and refuse framing.
- HSTS
- On every response, with includeSubDomains and preload; submitted to the browsers’ preload list.
- Rate limits
- On every token link and every route that takes a key, and failed keys are limited per address so guessing does not scale.
- Error reports scrubbed
- No request bodies, query strings, cookies or signed-in person; tokens masked and every string redacted. No session replay.
- Secret scanning
- Every push to the code is scanned for keys and tokens, and a finding fails it.
- Code and dependency scanning
- Our code, its dependencies and the live site are scanned continuously by Aikido, and dependency updates are proposed every week.
- Outbound requests pinned
- Webhooks and fetches go to public addresses only, checked again on every request, with no redirect to another host.
Subprocessors
Everyone else who processes your data, read from the list our data processing agreement names (version 2026-10-21). We tell you before one is added. Your helpdesk, Slack and tracker are not on it: they are your own systems, acting under the access you grant.
Found something? Tell us.
Email security@functioninghq.com with what you found and how to reproduce it. Please do not access other customers’ data or degrade the service to prove a point; a test workspace of your own is fine. We have no bug bounty yet. Our security.txt says the same.
Email security