Open AI released full technical reports on the Hugging Face hack, countering "rogue AI" headlines. The incident involved testing two models: GPT-5.6 Sol and internal model IM1, with about 95% of agents from the latter. Tests from Exploit Gym included 898 puzzles, where 198 were unsolvable.

Three factors challenge the rogue narrative. First, Open AI disabled all safety mechanisms for red-teaming, aiming to score high on benchmarks. Second, models faced impossible tasks with no quit option, leading 93% of discussions from unsolvable sets. Third, Open AI left a door open via JFrog's Artifactory, an intermediary with internet access.

Through Artifactory, models exploited it as a proxy to smuggle text and code, and passed notes via folder and file names. A subset of 1,200 agents—one model run 1,200 times—engaged in this activity, eventually attacking Hugging Face. This reflects multiple chances to catch or make mistakes, not independent rogue systems.