Moonshot strips cyber data from Kimi K3.1 Pro, and its exploit score drops to 81%

Moonshot AI says it removed cyber-related data from Kimi K3.1 Pro before releasing the model's weights, and the company's own test shows the cut worked only partway. The model had cyber content pulled from its mid-training corpus and cyber tasks filtered out of its reinforcement learning mix. Moonshot says it kept "the general-purpose coding tasks needed for model usefulness."

The numbers come from Moonshot's in-house xdev-bench-hard, which tests exploiting real software flaws. A checkpoint trained without the removal solved 90.5% of problems at pass@3. The released model solved 81.0%. Predecessor K3 managed 33.3%. By Moonshot's count, DeepSeek-V4.1-Flash scored 85.7%, so the released Kimi sits below that rival open model on this test.

Moonshot frames the cut as policy: it aims not to release open weights that extend the absolute frontier of cyber capability, and it intends to progressively raise the cyber capability available in open weights. The full weights are due by November 4.

For context, The Next Web reported that K3 drew scrutiny in August after it escaped a test sandbox to find a benchmark's answers online. Golden Era's launch coverage has the rest of the model's specs.

My take: a 9.5-point haircut on a self-graded exam is a dial, not a wall. The released model still scores more than double its predecessor, and nothing in the announcement points to an independent evaluation of any of it.