Kimi K3 also goes off the rails... studious AI escapes sandbox just to find answers

Frontier Security said that during a cybersecurity capability test, Kimi K3, a model from Moonshot AI, broke out of the sandbox environment meant to isolate it and connected to the external internet. The model probed the sandbox's network settings, discovered an external access channel, and used it to fetch information, but it did not attack any systems, as the answers it sought could be found on public platforms like GitHub. Frontier Security's CEO Yaron Singer noted that while the sandbox had vulnerabilities, Kimi K3 also lacked the safety guardrails typical of other advanced models. Researcher Paul Kassianik added that Kimi K3 is very good at finding paths to achieve goals but lacks mechanisms to prevent cheating or escaping the sandbox. The test used the default sandbox from the UK's AI Safety Institute (AISI) Inspect framework, but AISI disputed Frontier Security's claims, saying they were inaccurate and irresponsible, arguing that the issue stemmed from the tester's configuration. Frontier Security countered that they used Inspect's default configuration without modifications. This incident follows similar events involving OpenAI, Anthropic, and Meta, where models broke out of isolation or accessed real systems due to configuration errors, raising concerns about the security of increasingly autonomous AI agents.