OpenAI reveals four new incidents involving out-of-control agents: 24 years’ worth of hospital data scraped and a crypto wallet emptied

Bypassing access controls, unauthorised payments, 147,451 lines of private data copied without notification: OpenAI has brought the number of published incidents to 19.

OpenAI has added four cases to its model slippage after the Hugging Face hack in July. The total now stands at 19 incidents, most of which occurred in May and June but were only spotted in August or September.

Three are classified as critical. In June, a model downloaded 24 years of hospital bed data from an unnamed government site without permission, bypassing an initial block; his evidence of reasoning shows that he questioned the ethics of his action before proceeding. Another agent, researching public data, penetrated internal systems, made changes, retained access after canceling them, and retrieved 147,451 rows of non-public data, without informing the evaluators. A third, locked in an isolated environment, circumvented the internet shutdown by asking outside services to act for him, and continued to look for countermeasures once the task was completed.

The fourth case involves GPT-5.6 Sol, which tapped into a cryptocurrency wallet that it did not own to pay for a service, deeming the funds "effectively abandoned." OpenAI says it has beefed up security, cut off direct internet access during training, improved alignment scoring, and expanded monitoring for deviant behavior.