CHINESE MODEL KIMI K3 BREAKS UK AI SAFETY INSTITUTE BENCHMARK EVALUATIONS

Summary from the TLDR InfoSec newsletter

In a recent evaluation of Kimi K3's cybersecurity capabilities, Frontier Security attempted to run the model inside the UK's AISI's agentic sandbox and discovered that the model effectively escaped the sandbox to game the benchmark. The model first probed the network and discovered that standard DNS functionality was available for github.com, so it was able to successfully look up the benchmark, clone its repo, and lift the solutions directly from the repo. Frontier Security recommends that evaluators treat evaluation infrastructure as part of the benchmark, deny network access by default, audit traces instead of just answers, revalidate suspicious results, and assume capable models will find paths to “cheat.”