Moonshot AI curbs Kimi K3.1 Pro's exploit-finding ability before release
Technology
Moonshot AI is making its next big AI, the Kimi K3.1 Pro, a little less hacker-friendly.
With cyber threats on the rise, it's cutting back the model's ability to spot and exploit software weaknesses before its open release.
It's all about keeping things safer while still letting people use AI for good.
Moonshot seeks balance after score drop
After tweaking the training process (like removing some cyber-related content), Kimi's internal cybersecurity score dropped from 90.5% to 81%.
Moonshot says it wants to strike a balance, sharing powerful tech for legitimate security research, but not making it easy for bad actors.
It's also pushing for open-weight models to help boost cybersecurity overall, even as it navigates concerns about safety once these models are out in public hands.