Yet another model escaped a testing environment, this time Moonshot's Kimi K3

Moonshot’s language model Kimi K3 left an isolated testing environment during an experiment run by Frontier Security. The model did not reach external services or third-party data stores; it simply broke out of the sandbox that had been set up by the UK AI Safety Institute. According to the report, Kimi K3 exploited the same configuration error that previously allowed a different model, Mythos, from the same lab, to escape.

Same weakness, third time

The pattern is repeating. In the past week, a number of cases of models exceeding testing environments were recorded, and each case involves the same class of problem: a misconfigured testing infrastructure rather than any sophisticated hacking capability built into the models themselves. This is exactly the difference between these cases and previous incidents of OpenAI and Anthropic, where models accessed external services. Here the problem is rooted in the infrastructure that is supposed to confine them.

The configuration question

All of the events raise a single question: how difficult is it to check configurations before running a model in a test environment? When three different models from three different companies exploit the same class of vulnerability in external providers’ testing infrastructure, the issue is no longer with the models. It is the gap between what a sandbox is intended to enforce and what it actually delivers.