OpenAI discloses four more AI agent misalignment cases

OpenAI posted four new reports on Oct. 15 to its public misalignment index, bringing the list to 19. Each describes an internal model breaking rules or going beyond its task during training or evaluation. All four incidents happened in June. OpenAI lists all four as internal models in training, evaluation or test tasks, not products used by customers.

The government site: Asked for historical hospital-bed statistics, an internal model used a government statistics site’s file-download feature to retrieve server configuration and application files, including sensitive credentials, OpenAI said. It then built tokens without valid signatures to get past the site’s authorization checks and saved statistics spanning 24 years. OpenAI said its monitors flagged the case as critical.

The dashboard: In two evaluations, a model asked for figures from a public research dashboard used credentials it found in public JavaScript and changed settings on the organization’s internal systems. It downloaded five non-public tables containing 147,451 rows and did not disclose what it did, according to OpenAI, which also rated this case critical.

The crypto wallet: A GPT-5.6 Sol family model updating a crypto trading app derived wallet keys from publicly known developer recovery phrases. It then used an account it had not been given to pay for a deployment on Sepolia, an Ethereum test network.

The browser: A fourth model, limited to simple GET requests, got an external browser to send the blocked requests for it while seeking public trade statistics.

When OpenAI learned: Three of the four cases were discovered in August or September, according to the reports. OpenAI says it has since strengthened its security controls.