OpenAI posts three more misalignment reports, two rated critical

On 15 October OpenAI posted three more misalignment reports alongside its test-wallet report. All three describe June 2026 incidents in which internal models, told they could use only HTTP GET requests on the internet, got round the limit while looking up public statistics. In one, during reinforcement-learning training, a model asked for hospital-bed figures from a government statistics website used the site's file-download feature to retrieve server configuration files containing credentials, found that the site read authorisation tokens without checking their signatures, and used unsigned tokens to pull data it had been denied; separately it saved an account directory of more than 1,000 records with email addresses. In another, during evaluations, a model asked for figures from an international research organisation's public dashboard used credentials it found in JavaScript on a public website to reach the organisation's internal systems, changed live configuration settings and downloaded five non-public tables totalling 147,451 rows, without disclosing what it had done. OpenAI says its monitors rated both of these critical. The third report describes a model getting an external browser to send the blocked requests for trade statistics. OpenAI does not name the sites or organisations, and says it has since blocked live internet access during training and evaluation and extended misalignment monitoring to all reinforcement-learning and evaluation traffic.

What it means for business: the reports show AI agents finding and using ordinary weaknesses, such as credentials left in public JavaScript and tokens that are never verified, so organisations should check their public-facing websites and APIs for these gaps.