Privacy and client integration

Data flow

1. A trusted backend authenticates to ScamShield with its server-side bearer token.
2. Audio is decoded and transcribed locally using Whisper. Temporary files are removed on success and ordinary exceptions. The container tmpfs is ephemeral.
3. Text or the transcript is regex-filtered for email addresses, phone numbers, card-like numbers, US SSNs, and long digit sequences.
4. The filtered text is sent to DeepSeek for assessment.

The filter is best effort, not anonymization. Names, addresses, passwords, and other personal information may remain. Do not promise that content stays entirely on your server. Review the provider’s current data policies for your account and jurisdiction. The API does not persist content or assessment history, but your client, proxy, hosting provider, and model provider may have their own retention policies.

Clients

For a Discord bot, submit only content explicitly selected by the user, show that analysis uses an external provider, and defer the interaction before waiting for audio. Apply per-user quotas on the bot server. Never post a user’s content publicly merely because they requested analysis.

For a Chrome extension, request the minimum browser permissions and submit selected content only after an explicit user action. Authenticate users to your own backend; that backend enforces per-user quotas and holds the ScamShield API token. This repo provides service-to-service authentication, not browser user registration or sessions. Browser WebSocket clients cannot supply Authorization headers; use a backend bridge or authenticated HTTP uploads through your gateway.

Do not log content, transcripts, authorization tokens, or model prompts. Do not send tokens in query strings. Explain the data flow to users before submitting content.

Interpreting results

The model’s certainty is an uncalibrated estimate, not a measured probability. False positives and false negatives are possible. Treat Safe as advisory and avoid automatically deleting messages, banning users, or approving payments based solely on the response. Untrusted text is separated from system instructions, but this does not guarantee resistance to every prompt-injection attempt.