Edition
Report: Astra becomes a leading target for Chinese AI labs
The Information reports that DeepSeek and Moonshot AI used Astra-generated output extensively when training recent models.
The reported distillation depends on extracting hidden reasoning, a practice prohibited by both OpenAI and Anthropic. Anthropic says its stronger defences now sharply reduce the amount of data extracted. OpenAI says attackers have kept adjusting to new safeguards. Moonshot offered no comment; the report says DeepSeek did not respond.
This summary is based on the available full source text. Summary prepared with AI assistance. The original source is authoritative.
Original source
The Decoder
Date stated with the source: 29. 10. 2026
Open the original at the publisher ↗