Summary:
Anthropic said Monday that new measures aimed at preventing competitors from copying the capabilities of its Claude AI model have reduced the amount of data stolen by so-called "distillation attacks" by about 90% to 99%.
Model distillation has become a focus of top artificial intelligence companies; companies such as Anthropic, OpenAI, and Google believe that this practice allows competitors to copy their advanced capabilities at a lower cost and faster.
Anthropic said that safeguards aimed at limiting access to Claude’s inference capabilities have significantly reduced extraction attempts against the model, and that the “number of inference tokens extracted through distillation” has now dropped to “1/100 to 1/10 of the previous level.”
The California-based company said it has made it more difficult to leverage a network of agents to repeatedly probe Claude's protected inference process, blocking data extraction attempts and hampering large-scale model distillation campaigns.
Anthropic says its security system is "roughly equivalent to a system that would have prevented Claude from assisting in the creation of a biological weapon" and has a "false positive rate of less than 0.1 percent."
Comments