Scambaiter

general view
logo

Inspiration AI has made scams cheap. One scammer can send thousands of convincing messages a day, and blocking them costs the scammer nothing. We wanted to make scammers pay with the one thing they can't scale: their time.

What it does Scambaiter is an AI agent that lives in a burner email inbox. When a scam arrives, it replies as Margaret, a 78-year-old who is very interested but never manages to pay. While she stalls, it collects what the scammer reveals (crypto wallets, bank accounts, phone numbers, websites) and builds a report ready to file with the FTC or FBI. A dashboard shows the estimated time wasted and intel captured. Normal emails are ignored.

How we built it Python and FastAPI, with Qwen 2.5 72B through Featherless. Detection combines 16 explainable rules and link checks with an AI verdict. Each turn, the persona gets a stalling tactic and a goal for what to fish for next. A safety filter written in code checks every reply: no links, no numbers, never admits to being an AI. The app reads and answers a burner Gmail inbox over IMAP and SMTP.

Challenges we ran into Keeping the persona believable without ever leaking anything that looks like real data, which we solved with a code-level filter. Real scammers don't reply on schedule, so we built a second AI that plays the scammer for demos. Our first email version also skipped any message that had been opened in Gmail.

Accomplishments that we're proud of It works on real email: a scam lands in Gmail and the AI replies on its own within seconds. Every detection explains why it was flagged, and the output is evidence authorities can act on.

What we learned Building an AI agent is mostly about guardrails. The AI does the creative part, and ordinary code keeps it safe, in character and on goal.

What's next for Scambaiter A one-click forwarding address for anyone, shared intel across users so repeat wallets get prioritized, automatic reporting to wallet blocklists and registrars, and a voice persona for phone scams.