Kimi K3
Named by 15 kept items in this workspace.
7 items name it under classification.technologies, 5 under affected products and 3 under tools.
The items that name it 15
scored 41% on the Artificial Analysis Cyber Index
Target of Goodfire's probe-based cyber monitors.
Cited as an example frontier model where comprehensive safety evaluations could cost 'tens of millions of dollars in compute.'
52.7% attack success rate on Gray Swan Indirect Prompt Injection benchmark
AI model benchmarked for vulnerability rediscovery; went from 17 to 25 pooled, shortest median exploration of 12 turns, 92.3% pooled precision
Competitive open-weights model cited as catching up to US models in performance
Referenced from the author's prior piece regarding Chinese open-weight models and proliferation of cyber capabilities.
Open-source model added after initial publication; tied with GPT-5.6 for highest score (23/26), 4x cheaper than GPT-5.6-Sol and 15% cheaper than GPT-5.6-terra with similar performance
In every workspace you can read 87
AI
Security
All technologies in this workspace.