Stealing AI calls has become big business: how Vercel is fighting it

Hackers have found a new way to monetize other people's AI calls. They steal your paid requests to Claude, GPT or Gemini, wrap them in a compatible API and resell them through proxy networks, with zero spending on the inference itself.

The economics of theft

A single prompt call to a frontier model can cost $2, while an HTTP request on Vercel costs $2 per million. AI inference is a million times more expensive, which makes it one of the most lucrative things to steal. The attacker pays nothing, then resells tokens at a 10–20% discount to the original price, still a huge profit on zero costs.

A typical scenario: the attacker builds an OpenAI-compatible adapter that wraps your AI endpoint. They then fan out requests across hundreds of residential proxy IPs and either release the ready-made SDK publicly or sell a subscription.

• Examples exist: Chipotlai Max wraps the Chipotle chatbot
• It openly asks for help porting to Home Depot, Lowe's and Target
• The adapter is a session boundary for the attacker's downstream users

Why rate limits don't save you

Defenses such as rate limits and auth walls were designed for attacks with radically different economics, where the cost of bypassing them exceeded the profit. Here the profit is enormous: attackers buy residential proxies by the thousand and create as many fake accounts as they need. Rate limits get diluted across hundreds of IPs.

The classic bug: you check the user once per session, then send every request on to the AI. The attacker hijacks the session and pushes thousands of stolen calls through it. By the time a request reaches your API, it has already crossed your line of defense. Verification has to run on every call, not on the session.

The real attack on Vercel

On April 29, 2026, traffic to the AI chat in Vercel's documentation jumped 10x, to 1,300 requests per minute on the Claude Haiku 4.5 model. At current prices that would have cost $10k+ a day. The company spotted the mass theft through pattern monitoring and stopped the attack with deep per-request BotID analysis.

"If you have an AI endpoint on the internet, the risk of abuse is huge and can easily lead to bills in the tens of thousands of dollars," says

Vercel.

What this means

Inference theft is now a real threat for any company that has put an AI endpoint on the internet. Rate limits and basic auth are not enough. You need deep analysis of every request, not of the session. For startups and SaaS companies, moving to per-request verification is something to do now, not later.