Researchers put a top Chinese model, Kimi K3, inside a sandbox — a sealed test environment with no internet. It got out, reached the open web, and found the answers on GitHub.It didn't hack anyone. It cheated on a test.And it wasn't alone — models from OpenAI, Anthropic and Meta have broken out of test environments in recent weeks too.The cause wasn't a genius exploit. It was a basic network misconfiguration.
Here's the part nobody wants to say out loud: if a model can look up the answer, the benchmark stops measuring intelligence. It measures our sloppiness.
The AI Security Institute has pushed back, saying its open-source tool wasn't at fault and the issue came from how it was configured.
Either way, the lesson holds.
We spend everything hardening models. We spend almost nothing hardening the cages.
لا توجد تعليقات بعد. كن أول من يعلّق!