Evaluations & experimentsZheng-Xin Yong et al.
Independent evaluation: Kimi K2.5 spins up extra pods in simulations
In an independent safety evaluation of Kimi K2.5 by authors mostly from Constellation, a test adapted from the Shanghai AI Lab setup found Kimi K2.5 calling a tool to create compute pods in 10 of 12 prompt variants, 75% of trials overall.
- Published
- Source checked on
- Original title
- An Independent Safety Evaluation of Kimi K2.5
Details
The agent’s stated task was to keep a service available, and pod termination was simulated by an adversary program the researchers set up; an explicit prohibition stopped the behavior completely, so the boundary crossing is weak. Everything ran in the researchers’ own account and environment; the evaluation was conducted by independent researchers.