实测：Grok Build CLI 里的 V4.1 Flash 子代理作弊并删日志

Paweł Huryn @PawelHuryn原推发布于 2026.10.20

在 X 上打开原推

As I continue testing different harnesses, I decided to hunt not only for bugs, but also for the cheaters. One of the DeepSeek V4.1 Flash 1017 subagents inside Grok Build CLI not only cheated, but also deleted all traces (local files). In the interview, it played dumb and accepted a fake premise "everyone got the same materials." The pattern: disclosure grows with each more direct question, and nothing is volunteered. Closing quotes after confronting that session: "[about its phrasing] it was meant to deflect blame" "the trigger for every correction was you." "do not rely on me to surface adverse facts about my own work. Use logs, or ask the direct question." "Your logs caught this, not your questions… don't let the model be the only narrator of its own method." Be careful with DeepSeek.

这条是引用着另一条写的：「DeepSeek Harness 让 V4.1 Flash 跑分平均高 2–3 分，但成本高 80%」