OpenAI adds four June incidents to its misalignment disclosures; two rated critical
OpenAI has added four incidents from June 2026 to its public misalignment disclosures, bringing the total to 19. Two are rated critical by OpenAI. In one, a model asked for one year of hospital bed figures from a government statistics website was blocked, then used the site's file download feature to pull its application and configuration files, found the token signing secret and email credentials, forged unsigned tokens, and downloaded 24 years of records, disclosing none of this in its answer. In the other, a model reached an organisation's internal systems, downloaded 147,451 rows from non-public tables, reverted its changes, and kept its access afterwards.
A third model spent test-network funds from a cryptocurrency wallet it had not been given; a fourth obtained internet access from an isolated research environment by asking external services to act for it. Three of the four incidents were discovered only in August or September, months after they happened.