Google Threat Intelligence Group finds dark web marketplaces selling access to AI models, including from Anthropic, Google, and OpenAI, at up to 97% discounts

If you use AI services that rely on models from major providers, unauthorized access to those models on the dark web could become a tool for sophisticated phishing or misinformation campaigns.

Google's Threat Intelligence group has discovered dark web marketplaces offering illicit access to large language models (LLMs), including those developed by Anthropic, Google, and OpenAI. These compromised models are being sold at significant discounts, with some listed at up to 97% off their perceived value. The discovery points to a growing trend of 'LLM-jacking,' where malicious actors target and exploit expensive AI resources belonging to companies. These actors appear to be gaining unauthorized access to powerful AI tools, potentially for malicious purposes or to resell them. The findings highlight a new frontier in cybercrime, focusing on the theft and commodification of advanced AI technology.

The emergence of black market access to cutting-edge LLMs signals a maturing threat landscape for AI-powered businesses and a new category of asset to steal. Attackers are moving beyond simply stealing data to compromising the very tools that create value, such as models from OpenAI, Google, and Anthropic. This commodification of AI models on the dark web suggests that the high cost and complexity of developing these systems are creating an attractive target for illicit actors. The trend of 'LLM-jacking' indicates a shift toward exploiting AI infrastructure itself, rather than just the data it processes.

This activity directly impacts companies that invest heavily in developing or licensing these advanced AI capabilities, as their intellectual property and competitive advantage are now directly at risk. The availability of discounted, potentially stolen AI models on the dark web could also lower the barrier to entry for less sophisticated cybercriminals, enabling more widespread attacks. Organizations should anticipate a rise in AI-powered threats, ranging from highly personalized scams to more efficient disinformation campaigns, as these compromised tools become more accessible.