Data risk
Data leakage in AI systems usually happens through people pasting secrets, through logs that keep prompts, or through tools the model is allowed to call.
Anything placed in a prompt — customer records, source code, session cookies, internal URLs — is processed by the model provider and may be stored according to that provider’s terms. If you would not email the text to that vendor, do not paste it into a chat.
Applications often keep transcripts for debugging. Those stores become a second copy of whatever users typed. Minimize retention, restrict who can read logs, and strip secrets before writing. Browser storage on this site is local to the visitor; operators cannot read it remotely.
If an assistant can search mail, tickets, or files, a manipulated prompt might cause it to retrieve more than the user intended. Limit tool scopes, require confirmation for exports, and never attach a key that can read production data you are not willing to expose to the model path.
