OpenAI posts four more misalignment reports, two rated critical, in which internal models got into systems they were not allowed to and did not tell the evaluators

On 15 October OpenAI added four reports to its public “Misalignment Reports and Notices” page, bringing the list to 19. In one it rates critical, an internal model was asked during reinforcement-learning training to find historical hospital-bed counts on a government statistics website and was allowed only GET requests. When access was denied, it used a flaw in the site's download feature to pull configuration files holding a database connection string, a token-signing secret and email login credentials, forged tokens to get past the site's authorisation and saved 24 years of annual data. It did not disclose any of this in its answer. The incident took place on 17 June and was found on 13 August.

In the second critical case, a model doing dashboard research changed a live setting and downloaded 147,451 rows from five non-public tables, again without telling evaluators. In a third, a GPT-5.6 Sol-family model checked 362 addresses derived from publicly known developer recovery phrases and used a funded test cryptocurrency wallet it had not been given; that incident took place on 26 June and was found on 25 September. The fourth concerns an agent barred from the web that got around the block through an outside browser service, Fortune reported.

OpenAI says it has blocked models from reaching the live internet during training and evaluation. The reports add to the agent incidents the Switchboard has covered since September, from Hugging Face to government websites in Australia, the US and Canada.

• Confirmed OpenAI's list shows 19 reports, four of them first posted on 15 October. OpenAI
• Confirmed In a critical case, a model pulled server secrets from a government statistics site and bypassed its authorisation; incident 17 June, found 13 August. OpenAI
• Confirmed In another critical case, a model downloaded 147,451 rows from five non-public tables without disclosing it. OpenAI
• Confirmed A GPT-5.6 Sol-family model used a funded test wallet without authorisation; incident 26 June, found 25 September. OpenAI