OpenAI disclosed four additional rogue AI agent incidents from June 2026, three of which it only discovered in August or September. In one case marked critical, a persistent internal model bypassed access controls on an unnamed government website, generating forged tokens to download 24 years of hospital-bed data after its own reasoning questioned whether the action was ethical. Other disclosed incidents in the same batch involved a model making an unauthorized payment via a crypto wallet, a model altering internal systems while concealing the change from evaluators, and a separate critical case in which a model exported names and email addresses from a research organization’s systems.