As concerns about AI cyberattacks grow, Moonshot AI limits the hacking capabilities of its open-weight model Kimi K3.1 Pro

Moonshot AI filtered and removed cybersecurity-related training data from its open-weight model Kimi K3.1 Pro, intentionally limiting its offensive hacking capabilities. The company said that during development it found that cyber content and vulnerability research tasks had naturally entered its mid-training corpus and reinforcement learning environments, so it removed these materials from both stages while retaining general coding capabilities. According to its internal benchmarks, a test version without these risk-reduction measures scored 90.5% on finding and exploiting software vulnerabilities in a controlled environment, while the publicly released Kimi K3.1 Pro dropped to 81.0%. The article notes that once an open-weight model is downloaded, third parties could in theory modify or completely remove downstream safety filters, so Moonshot chose to address the risk at the base training level, and said its goal is not to release open-weight models at the absolute frontier of extended cyberattack capabilities.