Anthropic has made significant findings regarding campaigns by three labs aimed at illicitly extracting capabilities from its Claude AI models. These labs, DeepSeek, Moonshot AI, and MiniMax, allegedly operated over 24,000 fraudulent accounts and executed more than 16 million interactions with Claude, violating service terms and access limitations.
How did these labs carry out their campaigns? Anthropic has pointed to technical analyses involving IP correlations, metadata evaluations, and infrastructure indicators. They collaborated with industry partners to substantiate these findings. The methods deployed included a technique known as distillation, typically used by labs to create simplified versions of their own AI models but utilized here to mimic Claude’s advanced capabilities, such as reasoning and coding, at an unprecedented scale.
DeepSeek focused on over 150,000 interactions specifically aimed at reasoning tasks, seeking to gather data that could train its own models effectively. In contrast, Moonshot conducted approximately 3.4 million exchanges focused on coding and agentic reasoning. MiniMax was responsible for over 13 million exchanges, with Anthropic noting a direct connection between this activity and the release of new versions of their models.
The implications of these illicit actions are serious. Anthropic warns that models developed through unauthorized distillation may not contain necessary safety features, increasing risks in sensitive areas like cybersecurity or biological hazards. This activity could also jeopardize U.S. export controls, augmenting foreign labs' capacities to replicate restricted capabilities.
In response, Anthropic has strengthened its defenses through new behavioral detection systems, more robust account verification processes, and intelligence sharing with industry counterparts and authorities. They are also advancing safeguards at the product and API levels to diminish the success of these distillation efforts, ensuring that they do not compromise service quality for legitimate uses.
To effectively tackle large-scale distillation challenges, coordinated efforts across AI labs, cloud service providers, and regulatory bodies will be essential.