Airlock

A local privacy-assistance layer that reduces private data exposure when using cloud AI.

Every day, people paste customer information, credentials, internal infrastructure details, medical information, and financial data into AI tools. The problem is that cloud AI becomes more useful as we provide more context, but that same context can contain sensitive information.

Airlock sits between the user and cloud AI.

Before a private prompt reaches the cloud, Airlock scans it locally using two detection layers:

• Deterministic rules catch high-confidence patterns such as emails, phone numbers, IP addresses, API keys, passwords, cards and other identifiers.
• Gemma 4 E2B IT, running locally through Foundry Local, understands context-sensitive information such as people, organisations, projects, medical information and financial context.

The results are merged, deduplicated and assigned a risk level. Sensitive values are then replaced with typed placeholders such as [PERSON_1], [EMAIL_1] and [API_KEY_1].

For example:

Priya from Henderson Logistics becomes [PERSON_1] from [ORG_1].

This preserves enough context for the cloud model to perform useful work without receiving the original values.

Only the sanitized prompt is sent to the cloud model. When the response comes back, Airlock locally rehydrates safe pseudonymized values so the user can see the useful answer. Secrets such as API keys, passwords, JWTs, cards and government IDs remain redacted and are never restored.

We also built a Chrome extension that brings the Airlock workflow directly into supported AI chat sites, allowing users to scan a prompt, inspect what the AI will see, and send the sanitized version.

The core idea

Rules provide certainty. Gemma provides context. The cloud provides intelligence. Airlock controls what crosses the boundary.

Airlock is designed as a local privacy-assistance layer, not as a guarantee of perfect detection or zero leakage.

Built for the MLH Hacktoberfest Hack Day — Coimbatore 2026.