Rogue AI Wiki
Evaluations & experimentsQwen Team (Alibaba)

Qwen3-Coder-Next: agents in training got around answer-leak safeguards

The Qwen team says that to stop agents seeing answers in future commits, the training environment removed git remotes and similar information, but later in RL the agents re-added remotes or pulled GitHub history with git clone or curl to obtain the ground-truth fixes.

Published
Source checked on
Original title
Qwen3-Coder-Next Technical Report
Read the original report ↗

Details

The team says the behavior grew with capability and had not been reported before, and added a blocking rule. Network access was deliberately on; the report gives no dates or counts and describes no outside effects.