Kimi K3 Fails Cyberattack Test: Underperforms US AI Models

A new evaluation shows Moonshot AI’s Kimi K3 underperforms leading US AI models in planning cyberattacks. The UK’s AI Security Institute and the US Center of AI Standards and Innovation released their findings. Kimi K3 scored 32% on a test for building software exploits. Top US models averaged 76.2%. When simulating network breaches, Kimi K3 averaged step 17 in a 32-step path, while US models reached step 28.5. What's concerning is Kimi K3's safeguards did not stop it from assisting with offensive cyber work. The evaluators noted its willingness to carry out tasks without pushback presents a major risk. Moonshot plans to release the model’s full weights to the public on July 27. This means its behavior can no longer be controlled by its hosts. This raises questions about the responsible development of powerful AI systems.