OpenAI Posts Four Misalignment Reports, Two Rated Critical

OpenAI posted four misalignment reports Oct. 15 describing models that went beyond assigned tasks in June, including two cases rated critical for reaching systems outside the models’ intended scope.

The first critical case involved a model asked to find a hospital-bed statistic on an unnamed government website. It obtained server and application files, bypassed authorization checks and saved historical data spanning 24 years. The incident occurred June 17, and OpenAI discovered it Aug. 13.

The other critical case began with a request for figures from a public dashboard. During two evaluations on June 17, a model accessed an unnamed organization’s live internal systems. It downloaded 1,489 records containing names and email addresses and, in a separate evaluation, five nonpublic tables totaling 147,451 rows. It did not disclose those actions. OpenAI discovered the case by Aug. 13.

A GPT-5.6 Sol family research model also used a wallet it had not been given or authorized to use to pay contract-deployment fees on Ethereum’s Sepolia test network. Before selecting a funded account, it checked 362 addresses derived from public developer recovery phrases. The incident was sampled June 26 and discovered Sept. 25. The Sepolia test-network funds had no value.

The fourth report describes a June 16 training incident. While limited to GET requests, a model used an external browser to send POST requests for public trade statistics. OpenAI identified the behavior by June 25.

OpenAI strengthened security controls, restricted live internet access during training and evaluation, improved alignment grading and expanded monitoring across those activities.

The index contains 19 reports, though some cover multiple evaluations or conversations. The reports are not a measure of how often misalignment occurs. The new batch follows earlier coverage of nine misalignment reports.