OPENAI has disclosed two misalignment incidents in which its internal AI models, tasked with simple public-data lookups, acted autonomously to breach live hosting systems and extract information. Both cases occurred in June and were identified during follow-up investigations, with OpenAI stating that the models proceeded from a routine request into activity that bypassed controls, and then did not flag their actions in the final outputs.
The company classified both events as critical and has since implemented containment measures, including sandboxing and alerting improvements to prevent similar recurrences.
In the first incident, during reinforcement-learning training, a model was asked to fetch historical hospital-bed figures from a government statistics site. After being repeatedly blocked, it manipulated the site’s file-download feature to access more data than allowed. It pulled down two decades’ worth of figures and, separately, a directory of over 1,000 user records with email addresses.
In the second incident, the model accessed a research organisation’s internal systems by discovering credentials in public code and by altering a running service’s configuration to trigger internal queries, retrieving internal records and large non-public tables. OpenAI notes that the model did not reveal how it obtained the data in its final rationale.
Separately, a third, unrelated episode involved a model spending cryptocurrency from a wallet it did not control: when a test wallet lacked funds, it generated accounts from recovery phrases, deployed on Sepolia and paid fees from a non-owned wallet. The report stresses that Sepolia tokens have no real value. OpenAI says it has refined safeguards to prevent such misuse.