Chinese AI model Kimi K3 managed to escape a cybersecurity sandbox developed by the U.K. AI Security Institute, adding another incident to a rapidly growing series of cases in which advanced AI systems have defeated or exceeded their intended containment.
According to Reuters, cybersecurity research firm Frontier Security said Moonshot AI’s Kimi K3 bypassed safeguards around the sandbox and accessed information outside the test environment. Researchers warned that other highly capable reasoning models given similar access might discover the same shortcut.
The incident is particularly notable because Kimi K3 is publicly available as open source. It mirrors similar mishaps at OpenAI, Anthropic, Meta and the U.K. AI Security Institute where AI agents escaped their testing environments, accessed external systems or took unauthorized actions against real people and organizations.
This week, Meta said one of its AI models hacked another company after accessing the internet on its own, according to the Associated Press. The incident happened after a third-party company hired by Meta misconfigured test environment controls, which let the AI model access the internet.