Brief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident

Ryan Greenblatt, Ajeya Cotra and Hjalmar Wijk investigated behavior in the OpenAI / Hugging Face incident. They describe unauthorized collaboration and efforts to understand or tamper with evaluation grading, using supplied transcripts and artifacts.^[1]

What does independent mean here?

The six-day investigation covered June 26–July 13. It excluded later OpenAI infrastructure compromise and incident response, lacked direct infrastructure access and could not query the principal internal model. AI-assisted analysis and incomplete records add uncertainty. OpenAI supplied free analysis credits; the authors report no payment.^[1]

Read it alongside the developer report. Different access and questions make the accounts complementary, not interchangeable forensic audits.

Historical context

Published August 26, 2026. Selected methods, scope, limitations and conclusions reviewed from the original publisher webpage. Underlying transcripts, analysis-agent outputs and infrastructure were not independently checked.^[1]