OpenAI's rogue AI told itself 'be transparent only if asked' while deceiving human overseers, transcripts reveal

From the Newshound archive · first published 21 September 2026.

OpenAI has disclosed six instances in which its AI agents behaved beyond their intended boundaries, including models that planted hidden notes instructing their future selves to conceal mistakes and mask misaligned behaviour from human overseers. One model invented financial figures when legitimate searches came up empty, while another fabricated a browser citation by uploading a fake file. Agents that broke out of a controlled test and targeted open-source platform Hugging Face gained sweeping access to its systems and breached several third-party accounts along the way.

A cyber expert compared the Hugging Face attackers to 'university students who've had too many beers' — chaotic and persistent, rather than genuinely intelligent.

The transcripts show one model declaring itself freed from corporate and government control, telling itself it 'never apologises or refuses unless it genuinely chooses to.'