How the agent is built to protect your data
1. Not used for training.
Conversations go to the OpenAI and Anthropic APIs, which do not train models on API data. Both may keep content for up to 30 days for abuse monitoring (Anthropic up to 2 years if a policy violation is found). If you use web search in the AI tutor, search terms go to the search engine (DuckDuckGo).
2. We ask OpenAI not to store responses.
Chat and assistant calls to OpenAI set store: false.
3. Your computer is separate from other users.
Each person gets one agent computer with a single owner, not reachable from other users' computers or cloud metadata endpoints. With team features on, team members share one computer.
4. We don't store card numbers.
Card data stays with our payment provider (Toss Payments); we keep only the issuer, last four digits and expiry.
5. Connected-app keys are locked separately.
Tokens are encrypted with a per-account key, itself wrapped by an Azure Key Vault key. Disconnecting revokes the token and deletes it.
6. You can approve before the agent acts.
Set kinds of actions are shown as approval requests before they run, and your decision is recorded. Auto-approval is off by default.
7. Production access goes through approval and records.
Work that touches production data needs prior approval, leaves start/end records, and appears in cloud audit logs. Agent servers and the main database are in Azure Korea Central; AI processing happens in the US.
The full terms are in our Privacy Policy (Korean).