AI Inference Theft Defense Strategy

While HTTP request costs are extremely cheap, AI Inference costs for frontier models are very high, reaching several dollars per request. This economic gap gives attackers a highly profitable business model called Inference Theft, a threat that frequently occurs against real APIs.

Inference Theft refers to the act of using someone else's paid AI inference resources without authorization, or reselling them at a low price. Attackers bypass existing security systems in the following ways.

• Using large numbers of Residential Proxies to neutralize IP-based rate limiting.
• Creating numerous disposable accounts to get past authentication walls.
• Wrapping custom AI endpoints in OpenAI or Anthropic compatible adapters to make them usable from standard clients.

In particular, AI Playgrounds, where users can directly control prompts and model parameters, are the most dangerous targets. On the other hand, Support Bots with fixed system prompts are relatively less exposed, but risk factors still exist as attackers develop techniques to trick models into bypassing system prompts.

Existing session-based authentication or IP restrictions are insufficient to block the economics of these attacks. Vercel performs in-depth analysis on all AI requests through BotID, blocking these attacks through per-request verification rather than session-level verification.