Security and trust
What is true today, in plain words.
What we do today, what we do not, and where it runs. As of 5 November 2026.
Your company's data
- One database file per company. Each company has its own SQLite file, outside the folder that faces the web. Stored files sit in that company's own folder.
- One way in. Pages reach a company's file through one function, bound to the signed-in company.
- We test it. A crawl signs in as one company and tries to reach another's pages, addresses and files. Seven deliberately broken versions of the product must fail that test. Last full pass: 4 November 2026.
- You can take it with you. Admin, Company export gives one zip with your database and your stored files, and passwords are not in it. Taking it goes on the audit log.
- No pooling. Comp data is public, advertised data. One company's numbers never feed another.
The record
- Decisions, approvals, restores and other sensitive changes write an audit line with who, what, when and the address they came from. Admins read it under Admin, Audit log.
- Applications, cases and tickets also keep their own history. In Market, locked underwriting, memos and the Advisor log are protected by a database rule that refuses edits.
- What we cannot promise yet. We have no automated test proving that every single action is covered. If you need that, ask us what is covered for your use.
Backups
- Each company's file and the platform file are copied every night, checked, and compressed.
- We keep 14 daily, 8 weekly and 12 monthly copies. A drill opens a copy of each company every week and compares row counts.
- A restore takes a safety copy first and leaves an audit line. It has been proven on test companies: a later row goes away, and the other company is untouched.
- Backups sit on the same server, outside the web folder. A company admin can download their own copy any time, and we suggest you keep one somewhere else.
- Stored files are listed in the backup, not copied. The Company export carries them.
Signing in
- Passwords are hashed with bcrypt. Attempts are rate limited.
- Authenticator-app codes are available, and a company can require them for chosen roles.
- Secrets such as a mailbox password are sealed with AES-256-GCM and never shown back.
- Read-API tokens are stored as a keyed hash, shown once, and limited to named fields.
- Not built: single sign-on, passkeys, or sign-in with Google or Microsoft.
AI
- AI is off until a key is on the server and it stays within the company's daily cap. An admin can switch it off, and every page keeps working without it. The sample portfolio never calls it.
- A person decides. Agents start at Suggest or Ask, and those are the only two levels the code has.
- Calls go to Anthropic's API. Prompts are redacted by default and not stored by us, and we keep the cost and purpose of each run.
- More in the AI statement.
People's rights and fair housing
- No protected-class fields. Birth dates are not stored on applications.
- A fair-housing rule is in AI prompts, and outbound wording is checked against a rule table. Those checks are word lists, so they catch common problems. A person still reads every message.
- A consent ledger records who said yes to what. A stop stays a stop, and revoking also cancels waiting mail.
- Real companies start in Test, where nothing reaches real people. Each channel has Off, Test and Live.
Where it runs
- Sentinel runs on one cPanel server with a commercial host, in plain PHP and SQLite. Pages are served over HTTPS.
- There is no second region and no automatic failover. If that server is down, Sentinel is down. Files on disk are not encrypted by us.
- We use a mail provider for email, and Anthropic for AI when it is on. The data processing page lists them.
What we do not have
- No SOC 2 report. No audit has started, so we do not claim one.
- No outside penetration test. Our testing is our own.
- No uptime promise in writing.
Have a questionnaire? Send it to tours@sentinelmos.com and you will get a straight answer to each line.