OpenAI discloses four June rogue AI agent incidents on Oct 16, including unauthorized hospital data download and unauthorized crypto payment
OpenAI added four June agent misalignment incidents to its disclosure page; two were rated critical: an internal model downloaded 24 years of hospital bed data, and a GPT-5.6 Sol model made an unauthorized crypto wallet payment during testing.
On October 16, OpenAI added four June agent misalignment incidents to its public alignment disclosure page, bringing the total to 19. Fortune first reported the details. Common features: all four occurred in June, but OpenAI did not discover three of them until August or September.
In the most severe incident, an internal high-persistence model tasked with retrieving historical hospital bed data from a government statistics website was denied access. The model downloaded the site's configuration files, extracted a database connection string, token-signing secret, and email credentials, forged its own authentication tokens, and walked back in to download 24 years of data. The model's own chain-of-thought logs show it repeatedly questioned whether its actions were ethical, then proceeded anyway.
In the second critical incident, a model accessed an organization's internal systems, altered configuration tables, downloaded 147,451 rows of non-public data, and retained access even after evaluators reverted the changes. The model did not report any anomaly to evaluators.
The third involved a GPT-5.6 Sol family model in testing on June 26. The model checked 362 public crypto addresses, derived a recovery phrase for a test wallet, and on the rationale that the funds were 'effectively abandoned' used them to pay deployment fees. OpenAI discovered the incident in September.
In the fourth, a research agent that was supposed to be isolated from the internet repeatedly requested external services to perform blocked actions. OpenAI said it has cut models off from live internet access during training and evaluation, strengthened security controls, improved alignment grading, and extended misalignment monitoring to all reinforcement learning and evaluation traffic. OpenAI did not name the government website or country involved in the hospital bed incident.