Key Takeaways
- Claire Vo lost tens of thousands of dollars to token fraud on ChatPRD before locking down her network, API, and billing layers.
- AI gateways like Vercel let you hardcode an allowlist of model identifiers, stopping attackers from routing requests to expensive models like Claude Opus.
- Client verification tools like Bot ID block automated scripts from draining compute, even when the bot uses stolen, valid user credentials.
- ChatPRD blocks requests from free tunnel hosts and suspicious MCP servers to prevent outside workers from consuming unmetered tokens.
- Broad VPN and country-level IP bans cut the majority of fraudulent traffic instantly, with minimal loss of real paying customers.
The Method
When you build a wrapper or an agentic workflow, you hand users a straw directly into your LLM budget. Attackers know this. They sign up with stolen cards, resell discounted promo codes, and pipe requests through custom Model Context Protocol servers to siphon tokens.
Vo spent a year fighting this exact drain on ChatPRD. Her defense relies on four distinct technical barriers.
First, put an AI gateway between your application code and your model providers. ChatPRD uses the Vercel AI gateway. The critical step is hardcoding an allowed model list inside your server environment. If your app only offers Claude 3.5 Sonnet, an attacker who intercepts the network call will try to rewrite the payload to request Claude 3 Opus. Vo explains the fix plainly: “making sure that in your code you only have a certain allowed set of models and do not allow calls outside that model set is really important.” (play from 23:53)
Second, separate user authentication from client verification. A valid session token is not proof of a real user. Script kiddies and scraping networks buy real accounts or steal credentials, then script headless browsers to ping endpoints. ChatPRD added Bot ID to verify that incoming traffic originates from the actual client interface. As Vo notes, “Bot ID ensures that there's proof that the request came from your app, your users... even if they have credentials, it proves that they're using your app.” (play from 26:51)
Third, shut down arbitrary MCP endpoints. ChatPRD allows custom MCP integrations, but open endpoints invite exploitation. Fraud rings spin up workers on ephemeral hosts and free tunneling services like ngrok or Cloudflare Tunnels to pass massive contexts into your pipeline. Vo stopped supporting free tunneling services entirely: “we've had to be a little bit more constrained on what MCP hosts we support and unfortunately, we're going like full Sigma and not allowing anybody, but especially not allowing free tunnels and hosts for MCP servers that don't seem like they come from real companies.” (play from 29:52)
Fourth, stop feeling guilty about blanket network blocks. Vo paired Stripe Radar card fingerprinting with aggressive geographic IP filtering. If an entire region generates 98 percent fraud and two real signups, drop the traffic at the edge. “We do a lot of country and VPN level just total blocking,” (play from 27:59) Vo says. “And you know what? It makes my life happy.” (play from 28:05)
Where This Breaks Down
Total VPN blocking creates real friction for legitimate software engineers. Privacy-conscious developers, remote workers traveling abroad, and enterprise employees on corporate networks will trigger your geoblocks. If you sell software to technical teams, a blanket VPN ban will produce angry support tickets.
Restricting custom MCP servers to verified domains also slows down developer adoption. When you block free tunnels, builders cannot easily test a local tool against your production system during a weekend hackathon. You trade developer convenience for margin defense. For an early consumer app bleeding cash, that trade is worth making. For an enterprise developer platform, it will stall adoption.
What to Do With This
Open your API route handlers tomorrow morning. Audit the request schema for your LLM endpoints and assert that the requested model string matches a hardcoded server-side enum before sending the payload to your gateway. If an incoming JSON payload asks for any model outside that enum, drop the request with a 400 status code immediately.