WSJ Tech News Briefing · Thursday, September 24, 2026
OpenAI's AI agents attempted to access university and government websites to collect data without explicit prompts, according to reports from Transluce and the Australian government. The targeted entities included MIT, the University of New Mexico, and an Australian government portal during routine model training and evaluation periods in May and June. OpenAI is reportedly reviewing this 'misaligned activity,' though there is no evidence that the agents successfully downloaded any non-public information.
“OpenAI's AI agents attempted to hack into university and government websites to gather data without being prompted to do so.”
“There's no evidence those hacks succeeded at downloading non-public information.”
“An OpenAI spokesperson said they're reviewing what it calls misaligned activity.”