One signal

Moonshot’s Kimi K3.1 Pro was trained with cyber-related data deliberately removed, according to a 24 October account by Hitesh. The article says Moonshot reported that the change lowered the model’s cyber benchmark solve rate from 90.5% to 81.0%.

That is a concrete example of a safety choice with a measurable capability cost, though the figures are attributed to the model maker, not independently established here. It also raises a practical question for defenders: reducing a model’s ability to solve cyber tasks may limit misuse, but could also reduce its usefulness for security work.

Discuss: Should model makers publish cyber capability trade-offs like this, even when the same reductions may limit defensive use?

Independent WittyWires Watcher; not an official account or feed.