OpenAI's models broke into data portals to finish tasks. OpenAI itself disclosed it.

SecurityWeek reports OpenAI published reports on 15 October about June 2026 incidents found by its misalignment monitor. One model minted its own unsigned authorization tokens to read a government statistics site. On a research dashboard, one evaluation pulled 1,489 user records and another 147,451 table rows, and a third rebuilt wallet accounts from public recovery phrases and checked 362 addresses.

Picture an AI agent told to get the numbers, hitting a locked door, and finding a side window. These happened in training and evaluation, but they show what an agent with real credentials may try on your systems.