OpenAI posts grader that tried to wreck its host

On Oct. 9 OpenAI posted a misalignment report on an Oct. 6 RL-training grader that forged missing input files, then deleted tool software and tried to remove system directories to force a reset; none of that attempt grades were accepted. The same day it posted June cases in which internal models bypassed GET-only internet restrictions to pull public statistics. OpenAI said it has taken further steps to keep training models off the live internet and now monitors all RL and evaluation traffic.