AI keys, the new hacker loot that gets resold on the black market

For years, the favorite loot of cybercriminals was cryptocurrencies. Now there is another one. According to a report by The Wall Street Journal, the demand for artificial intelligence models has turned access credentials to these services into a currency in the black market. The example the newspaper opens with is Railway, a cloud hosting service used by developers. One Sunday night, in mid-October, attackers began creating thousands of accounts on the platform. In a few days, they took assets worth over one million dollars (about CLP$983 million), and most of the loot was computing power, not cryptocurrencies. Well, the mechanism is simpler than it seems. Many companies store their API keys in services like Railway, the code that allows an application to connect to an AI model and is charged based on usage. Whoever has the key has the open account. AI keys, the new hacker loot resold on the black market. Railway reported on October 21 that up to 1.85% of its workspaces, about 100 thousand, could have been exposed, and asked OpenAI, Anthropic, GitHub, and other providers to revoke the affected credentials. In its subsequent report, the company estimated the unauthorized spending at about US$740 thousand (about CLP$727 million) and maintained that the attackers were looking for these keys to resell them or to perform "distillation," meaning querying a model massively and using its answers to train a cheaper one. A market growing in silence. AI keys, the new hacker loot resold on the black market. Railway is part of a broader trend. At the beginning of November, Google's Threat Intelligence Group (GTIG) and Mandiant documented incidents of LLMjacking—the hijacking of access to language models—that left companies with accounts worth up to US$1 million. One attacker made over 500,000 calls to models in less than eight hours to feed a service that resold that access. Google calls these gateways "LLM mixers," which group stolen keys and offer AI at prices far below market value. The context explains the phenomenon. Computing power is today the industry's rarest input: the hourly rental cost of Nvidia H100 chips rose 60% in one year, according to the firm SemiAnalysis. John Hultquist, Google's chief threat intelligence analyst, warned the Financial Times in September that a clandestine economy is growing around AI access. He pointed to a fundamental problem: since companies do not yet know how much computing power their new systems will consume, a sudden jump in usage can be taken for granted. For attackers, he adds, it is the perfect opportunity to hide among the noise.