Buying cheap AI tokens could expose your data to China.  Find out why scams are on the rise

Some users who are looking for a break on these prices may find sites on the “gray” market that charge much less, often discounting the price to between 70% and 90% off the normal cost of access.

While this may seem like a good deal, the reality is that these resellers are typically behaving illegally, and using the services they offer could put you at serious risk.

AI token fraud is on the rise

Okta researchers sounded the alarm on AI token fraud recently, indicating that they had found more than half a dozen services advertising discounted AI tokens, including sites like Poison Claude. That particular site alone had over 870 users, as a flaw in its configuration revealed.

While there are slight variations in exactly how the process works, typically Poison Claude, and other fraudsters like it, use nefarious means to gain access to Anthropic (which created Claude) and other AI models. And they have a number of tactics to do this.

“The primary attack vectors are free-trial abuse at scale, stolen or prepaid payment cards, chargebacks filed after the tokens have already been consumed, and proxying traffic through public inference endpoints that lack sufficient access controls,” Matt Lenhard, founder of Vectoral, a company that fights token fraud, told Moneywise.

In other words, the scammers may create many accounts to sign up for introductory trials with Claude to get access to credits for tokens, or may pay for tokens, but then get their credit card company to reverse the charges.

Scammers also exploit publicly available AI inference endpoints (like AI chatbots companies use for customer service) if the bot lacks proper authentication or usage controls. This allows the scammers to route requests through those systems without paying for the underlying AI use.

The companies add all of the free tokens they get from these tactics to a pool, then charge customers either subscription fees or flat fees to access the AI. The fraudsters give users an API key that allows them to send prompts to Claude and other large language models (LLMs), with the traffic routed through the scammer’s sites.

These services are popular both with people looking for discounted AI and also with users in countries like China where access to frontier US AI models is banned or blocked. While OKTA explained that AI companies have tried to stop this by taking steps like limiting free trials, the gray market is still thriving, and fraud is only increasing.

“The problem is growing rapidly,” Lenhard said. “I’ve spoken with companies where the total cost of this type of abuse has been doubling roughly every two months, and others that are losing more than $1 million per day.”

Companies like OKTA and Lenhard’s Vectoral are developing new and innovative ways to fight the fraud, but this is an industry in its infancy, and many companies that are eager to embrace AI service aren’t even aware they’re at risk or that LLMjacking even exists until they’ve already become victims.

Your data could be the price you pay

AI companies are taking a huge hit from this fraud, especially as Fortune reports that Stripe’s CEO, Patrick Collison, said as many as one in six new customer signups for AI firms are signing up only to resell the tokens from the free trials.