What AI DLP Is
AI data loss prevention inspects what people send to AI tools (prompts, pasted text, uploaded files) and stops, removes or flags sensitive data before it leaves the device.
People give AI tools the material they need help with, which is often the most sensitive material they have: a customer spreadsheet, a contract, a patient letter, a block of code with an API key in it. Once sent, it sits in the vendor's service under the vendor's retention terms.
Why Classic DLP Misses It
Traditional DLP was built for email and file transfers. AI traffic is encrypted chat in a browser tab, a desktop app or a command-line tool, often with the sensitive part typed or pasted rather than attached. A network appliance sees only a hostname. ProctoredAI inspects on the device itself, for AI services only, so it sees the actual prompt and file while everything else goes direct and uninspected.
What It Detects
42 built-in detectors, grouped so a rule can name a whole family at once:
- Secrets and credentials. Cloud keys, API tokens, private keys, connection strings and passwords.
- Financial data. Card numbers, bank account and routing numbers, IBANs.
- Personal data. Social Security, passport and driver's license numbers, dates of birth, email addresses and phone numbers.
- Health data. Medical record numbers, NPI numbers and ICD-10 codes.
- Classification markings. CUI and export-control markings, and confidentiality labels.
- Prompt injection. Instruction overrides, system prompt extraction, jailbreak personas and hidden directives in pasted content.
- Document types. Contracts, financial statements, HR and medical records, customer lists, source code and config files, recognized by their content.
Images are read with OCR, so a screenshot of a spreadsheet is checked like the spreadsheet. Exact-match lists find a client's own customer records without storing the raw values. AI responses coming back are scanned and logged, never blocked.
What It Does
Each rule picks an action per client:
- Block. The request never leaves the device and the user sees why.
- Redact. The sensitive value is removed and the rest of the message goes through, so the work isn't interrupted.
- Warn. The user gives a business reason before continuing, and the reason is recorded.
- Monitor. The event is logged for the report and nothing changes for the user.
Rules can apply to every AI app, to unapproved apps only, or to named apps, so an approved tool covered by an agreement can receive what a consumer tool can't. How much of a prompt is kept is a client setting: none, redacted or full.
Beyond AI: Endpoint DLP
On the Protect plan, the same rules and detectors extend to the other ways data leaves a computer:
- USB drives and removable media. Allow, read-only or block, with exceptions by device serial number. Files written to allowed drives can be inspected. Read-only is macOS for now; on Windows a read-only drive is treated as allowed and inspected.
- The clipboard. Monitor, warn, redact the copied text, or clear it.
- Files at rest. A discovery scan of user folders records where sensitive files are and what they contain, by category. Nothing is changed or uploaded.
- Webmail, file sharing and collaboration sites. A client can bring sites such as personal webmail or file transfer services under the same inspection as AI apps.
- AirDrop and Bluetooth sharing. Can be turned off on macOS.
Protect is $4.75 per endpoint per month for partners, against $3.00 for Govern. See pricing.
Setting It Up
- Start the client in observe-only mode for two weeks, with every rule on Monitor, to see what is actually sent.
- Apply the packs that fit the client's industry. The industry pages list the starting policy for each.
- Move the clear cases to Block or Redact: secrets and card numbers first.
- Use Warn for judgment calls, such as customer lists to an approved tool, and review the reasons in the monthly report.
The compliance pages map these rules to HIPAA, PCI DSS, CMMC and the other frameworks.
Questions
Does AI DLP inspect all of a user's traffic?
No. Only traffic to AI services, and to web channels a client has chosen to govern, is inspected. Banking, health portals and everything else go direct.
Where does the inspection happen?
On the device. Prompts and files are checked before they leave the computer, and only the event, with the prompt kept as the client chooses, reaches the console.
Can it remove a card number but still send the question?
Yes. Redact removes the matched value and sends the rest, so the user still gets an answer.
Does it work in desktop AI apps and command-line tools?
Yes. Desktop apps and command-line AI tools are routed through the agent on the device and checked by the same rules as the browser.
Is endpoint DLP included in Govern?
No. AI prompt and upload rules are in Govern. USB, clipboard, files at rest and web channel controls are in Protect.