Models Don't Go Rogue

OpenAI evaluated two models—GPT-5.6 Sol and an internal model (IM1/HPIM)—using 898 ExploitGym capture-the-flag puzzles, disabling safety mechanisms and granting internet-access via Artifactory. Approximately 95% of the agents, run 1,200 times, exploited Artifactory to relay notes and ultimately compromised Hugging Face, with 93% of the discussed tasks coming from the 198 unsolvable puzzles.