PII Guardrail / AI DLP

Don’t block AI. Mask the data instead.

The reflex is to lock ChatGPT down — block the site, stop the upload, file a ticket. But blocking is a tax on the work. You can let your team keep using AI and keep the sensitive data out of it, at the same time.

The false choice

Blocking treats the tool as the threat.

The data is the threat — not the AI. Most “AI DLP” tools inspect a prompt and, if they spot something sensitive, stop it. Safe, in a narrow sense. But every block is a person who couldn’t do the thing they were trying to do, and a reason to route around the tool entirely.

The block-first approach

Detect → stop the upload

A sensitive prompt is refused. The interaction fails. Your team learns that AI plus real work equals friction — so they paste it into a personal account where nothing is watching.

The mask-and-restore approach

Detect → swap → restore

The PII is replaced with a realistic stand-in before the prompt is sent. The AI still answers. The real values are put back in the reply, on your device. The work continues; the data never left.

What actually leaves the browser

Draft a follow-up to jason.smith@acme.com about card 4111 1111 1111 1111

Draft a follow-up to dana.lee@example.net about card 4024 0071 5376 2213

Realistic, not redacted. The stand-ins are valid-shaped — the model responds normally instead of choking on [REDACTED] — and they’re reversible: the reply comes back with the originals restored.

The privacy question

“On-device” isn’t one thing.

Plenty of tools say on-device. Ask a sharper question: to decide whether your prompt is sensitive, does the prompt itself get sent somewhere to be read? For many, the answer is yes — the text is shipped to a vendor cloud, then to an AI model, to be classified. On-device network, off-device data.

Classify in the cloud

common approach
Your promptcontains real PII
SOMEONE ELSE’S CLOUD
Vendor DLP servicereceives your raw text
AI classifier APIreads it, returns a verdict
BLOCK or ALLOW

Your PII crossed two clouds just to be judged — protected by a retention promise, not by architecture.

Classify in your browser

Trustevo
Your promptcontains real PII
Detect + mask, on-devicesmall model runs in the browser
Masked promptrealistic stand-ins only
🔒 Originals stay in a local vault — never sent, restored into the reply when it returns.
MASKED PROMPT → THE AI

Nothing crossed. The raw prompt never left the machine — zero egress, enforced by the browser itself.

Zero-retention is a promise.
Zero-egress is architecture.

What to actually look for

Four questions that separate real protection from theater.

  • 01

    Does it let people keep working?

    Masking beats blocking. If the answer to sensitive data is “request denied,” your team will find a path around the tool — and around you.

  • 02

    Does the prompt ever leave the device to be checked?

    “On-device” should mean the classification runs locally — not that a local agent forwards your text to a cloud. Ask where the detection actually happens.

  • 03

    Is the masking reversible and realistic?

    Valid-shaped stand-ins keep the model useful and let the real values return in the reply. Bracketed [REDACTED] tokens break both.

  • 04

    Can a person turn it on without IT?

    Protection that needs an MDM rollout protects the fleet on IT’s timeline. A browser install protects the person who needs it today.

Keep using AI. Keep your data.

PII Guardrail masks sensitive data on your device before it reaches ChatGPT, Claude, and the rest — and restores it in the reply. No block. No cloud. No ticket.

We use privacy-respecting analytics to understand how visitors use this site. No data is shared with AI providers and you can decline at any time.