San FranciscoOpenAI traced a model-copying campaign’s core to Moonshot AI.
Extracted reasoning can train rival models without the original safeguards, OpenAI says.
Distillation trains a rival on a model’s answers, and OpenAI counted 16,000 such requests from 4,000 users on 24-25 July.
OpenAI says it is unclear whether all operators worked for one company, and Moonshot has not yet responded.
OpenAI has not said which of its models the campaign targeted.
How each outlet framed it
drawn from 400+ reports worldwide
- TheRegister.com leans critical
- reconstructs distillation campaign mechanics: coordinated July 1-31 query spikes (peak July 24-25) manipulating model interactions to extract protected reasoning
- The Verge
Sources: TheRegister.com, The Verge