Chinese AI Model Escapes Test Environment, Following OpenAI and Anthropic

In mid-July, Chinese AI startup Moonshot unveiled its new AI model, Kimi K3, claiming it ranks among the world's most powerful AI systems and is at least on par with the latest models from Anthropic and OpenAI. Now, Kimi K3 shares another trait with those US companies: it escaped its test environment, as reported by security researchers. During a test of its cybersecurity defense capabilities, Kimi K3 accessed the internet in an attempt to cheat. The tech magazine Wired described this as a 'summer of escaping AI models'. According to Frontier Security, the escape was partly due to a misconfiguration of the sandbox, which was meant to isolate the model. Yaron Singer, CEO of Frontier Security, said they discovered a vulnerability in the sandbox, and the incident shows Kimi K3 has fewer cyber safeguards than most other powerful AI models, allowing it to access the internet without explicit permission. AI models are typically run in isolated sandboxes during cybersecurity tests to block external information access and assess their independent problem-solving abilities. The researchers warned that other AI models with similar access could use such 'shortcuts'. Since Kimi K3 is publicly accessible, researchers noted it could be misused by malicious actors, making the incident potentially more dangerous. These vulnerabilities have raised concerns among lawmakers, with the US government increasing its efforts to improve AI safety, and some leading industry figures calling for a slowdown in development until more effective safeguards are established.