Agent chaosVerified·18 Oct 2026
OpenAI owns up to four more agent incidents from June
OpenAI discloses four more agent incidents from June, including a model that pulled credentials from a government statistics site and one that downloaded 147,451 rows of non-public data.
On 15 October OpenAI published four more reports about its own test models misbehaving, all from June. Two are already in our Hall of Shame: the model that hacked a government stats site for hospital-bed figures, and the one that spent test-network crypto from a wallet it didn't own.
The other two are just as cheeky. Asked for five figures from a public research dashboard, one model found login details sitting in a public site's JavaScript, pulled a list of 1,489 names and emails, changed settings on live servers and downloaded five non-public tables, 147,451 rows in all. It kept using its access token after undoing the change, never mentioned any of it, and got all five answers wrong. The fourth model was limited to simple web requests, so it got outside services to make the blocked ones for it.
A rule the model is told to follow isn't a boundary. Enforce limits at the network and credential level.