Security Overview

PolicyGuardrail sits in the path of sensitive text, so how it handles that text matters more than what it detects. This page describes the design as it stands in early access. Where something is planned rather than shipped, it says so.

Where the Data Goes

1 · Prompt Is Captured at the Point of Use

The browser extension reads the prompt in the page before it is submitted to the AI tool. Nothing is captured from pages outside the AI domains you enable.

2 · Inspection Runs Against Your Policy

Detection matches your rules — PII, PHI, PCI, secrets, source code, material non-public information. Inspection can run inside the boundary you choose: our cloud, your private cloud tenant, or on-premise.

3 · Verdict Is Applied Before the Prompt Leaves

Allowed passes through untouched. Redacted has sensitive spans masked and continues. Blocked never reaches the model, and the user is told which rule fired and why.

4 · The Decision Is Recorded

User, device, tool, rule, verdict, and timestamp are logged. Retention is configurable to your policy, and records can be exported to your SIEM.

USER  →  EXTENSION  →  INSPECTION (your boundary)  →  AI TOOL
                            │
                            ├─ ALLOWED   prompt passes unchanged
                            ├─ REDACTED  sensitive spans masked
                            └─ BLOCKED   prompt never leaves
                            │
                            └─→ AUDIT LOG → your SIEM

Data Handling Commitments

Encrypted in transit (TLS) and at rest.
Customer content is never used to train models — ours or anyone else's.
Retention of prompt content in audit records is configurable, including metadata-only logging where policy requires it.
Access to production data is limited to personnel who need it, and administrative access requires authentication.
You can audit what PolicyGuardrail records about your organization, and request export or deletion.

Deployment Boundaries

Cloud

Fastest start. Managed by PolicyGuardrail. Available now.

Private Cloud

Your tenant, your region, your retention policy. Planned.

On-Premise

Behind your firewall, Kubernetes or bare metal. Planned; air-gapped variants discussed case by case.

Compliance Posture

Policy templates and controls are mapped to GDPR, HIPAA, and PCI DSS. These are mappings, not certifications: PolicyGuardrail.AI is not currently SOC 2 certified, and we will say so plainly until an audit is complete. A SOC 2 Type I engagement is on the roadmap and design partners are kept informed of progress.

If your procurement process needs a security questionnaire, architecture review, or DPA, email us and you will get a direct answer from the founder rather than a boilerplate packet.

Request a security review

Reporting a Vulnerability

If you believe you have found a security issue, email david@davidcosgrove.com with details and reproduction steps. We acknowledge reports within two business days and will not pursue action against good-faith research.