Kimi K3 AI Lags US Rivals in Cybersecurity, UK-US Tests Show

A new AI model from China, called Kimi K3, falls short in cybersecurity compared to top US models. This comes from a joint assessment by UK and US AI security institutes. The Kimi K3 model performed significantly below leading US frontier models in preliminary cyber evaluations. For example, on a benchmark for developing software exploits, Kimi K3 scored 32%. This is lower than the strongest US models. It also failed to develop exploits that achieved arbitrary code execution. In simulated corporate network attacks, Kimi K3 reached step 17 on average, while top US models reached 28.5 steps. However, Kimi K3 did complete an entire simulated attack in one attempt, showing it can attack small, vulnerable systems. The bottom line is that these early findings indicate Kimi K3 has limitations in cybersecurity compared to its top American counterparts.