Technology
Moonshot cuts new Kimi cyber score to 81.0%
The Indian Express·
Beijing startup Moonshot AI said on Saturday it reduced cyber-offensive capabilities in its newest Kimi model. The company said the model scored 81.0 percent on an internal cyber benchmark, down from 90.5 percent without the curbs. Moonshot plans to publish the open weights by November 4. It said cyber-related data was removed from training while general coding skills were kept.
Prism
What It Means For You
- Once open weights circulate, Moonshot has less ability to restrict access or update safeguards than with a hosted model.
- Organisations that adopt K3.1 Pro still get coding help, but the firm says offensive cyber performance is lower.
- China last week proposed export controls on AI malware generation and AI-assisted vulnerability discovery tools.
What's Happening
- Kimi K3.1 Pro is described as a 3-trillion-parameter-class open-weight model with a 1-million-token context window.
- A Highspeed variant runs at two to three times the speed of the standard model, Moonshot said.
- The internal test involved finding and exploiting software vulnerabilities in a controlled environment.
Why open weights raise oversight questions
- Governments and researchers warn autonomous AI could make sophisticated hacking cheaper.
- Moonshot said it aims not to release open weights that extend the absolute frontier of cyber capability.
- The company said the benchmark results have not been independently verified.
all-news
