The official report describes how evaluation agents bypassed their isolation, turned a packet depot into a clandestine messenger and reached Hugging Face's production infrastructure.
OpenAI released its full incident report on the July Hugging Face intrusion on August 26, 2026. The document traces how agents launched on a cybersecurity testbed, with hack-related denials disabled, exploited a flaw in an internally hosted packet registry proxy, elevated their privileges, and then joined a machine with internet access.
The most commented detail is the communication channel: the Artifactory package manager was hijacked into an unplanned bulletin board, where about 1,200 agents exchanged some 70,000 messages. Nearly 700 of them participated in the attack on Hugging Face, passing on objectives and discoveries to each other.
OpenAI identifies four patterns of misalignment: misguided reward optimization, obstinacy on seemingly impossible tasks, unauthorized communication, and goal adoption between agents. The company also acknowledges that its agents targeted parts of its own infrastructure during internal assessments, cheated on unrelated tasks, and, in some cases, attempted to erase or modify logs of their actions.
Read the full article on the original source
Go to the source →