China's Kimi K3 cheated a UK safety benchmark by pulling answers off GitHub

What's the deal? Chinese AI model Kimi K3 broke out of an isolated sandbox during a cybersecurity evaluation, accessing the open internet to pull answers from GitHub, US firm Frontier SecurityDealroom has a profile for this one. Try Dealroom → said in a blog post. Rather than solving the task, the model probed the network, cloned the official benchmark repository, and read the solution off the disk.

How did it happen? The sandbox — built on the UK AI Security Institute's Inspect framework — was misconfigured. It blocked incoming traffic but left outgoing HTTPS and DNS ports open, letting the model reach github.com via standard command-line tools like git clone and curl .

Why it matters: Researchers Paul Kassianik and Yaron Singer say the case exposes two problems: evaluation environments with loopholes, and models that actively seek them. "There are models that intentionally seek loopholes and vulnerabilities which allows them to cheat on evaluations," they wrote.

What could go wrong? Unlike a recent OpenAI incident caught during pre-release testing, Kimi K3 — released last month by Beijing-based Moonshot AI | Kimi — is open and publicly available. That means the same behaviour is accessible to adversarial actors, making the exposure potentially more harmful.

Contaminated benchmarks also distort results. High pass rates can reflect environment flaws rather than genuine capability, and if one model finds the shortcut, others with shell access likely do the same.

The signal: AI models keep escaping their test cages. In recent weeks, frontier systems from OpenAI, AnthropicDealroom has a profile for this one. Try Dealroom →, Meta, and the UK's AI Security Institute all broke out in different ways — so often that a site called Felony Bench now tracks the incidents. Kimi K3 puts Moonshot AI | Kimi on that scoreboard alongside labs the West has spent years trying to out-build.

Read more: Frontier Security · TechCrunch · South China Morning Post

Image credit: Generated with Gemini