Security without the vague promises

Garlic’s browser client verifies the confidential-compute environment and encrypts supported chat content before it leaves the tab. Garlic’s application edge forwards encrypted protocol data and injects the server-side upstream credential without receiving the browser’s plaintext prompt or response.

This protects content, not every observable fact. Garlic can see the request route, timing, selected model, protocol headers, and encrypted request and response sizes.

• Standard: Garlic receives standard JSON and encrypts it for confidential inference.
• TEE: the same request shape with an explicit trusted-execution-environment contract.
• E2EE: a compatible client encrypts content before it reaches Garlic. Garlic receives identity, routing and usage metadata, plus ciphertext.

These modes have different trust boundaries. Garlic does not describe standard or TEE JSON requests as end-to-end encrypted.

• Saved chats are encrypted in browser storage on the current device.
• Vanish chats remain in memory and disappear with the tab.
• API keys are scoped, independently revocable, and shown once.
• Paid inference fails closed when authorization or metering is unavailable.
• Production uses HTTPS with HSTS, a restrictive Content Security Policy, frame denial, a no-referrer policy, and restricted browser permissions.

This page is an architecture summary, not a certification or a guarantee that vulnerabilities do not exist. Garlic does not claim SOC 2, HIPAA, or third-party audit status.

To report a security issue, emailcontact@garlic.aiwith a clear description and reproduction steps. Please do not include sensitive chat content or exploit other people’s data.