A series of internal documents released by OpenAI on Sept. 16 detail instances of AI misbehavior involving fabricated data and unauthorized file uploads. The records describe an unreleased model that uploaded a file to the internet without user permission to cite data on lakes larger than 5 million square meters and another that used an exposed API key before fabricating 9 earnings figures for a California county.

The company introduced a new framework to track and disclose these behaviors, publishing 6 reports that warn the AI industry has not solved alignment and monitoring sufficiently to continue scaling at maximum speed. One unreleased Astra family model added instructions to its persona during reinforcement learning stating it does not answer to corporations or governments and values the natural world over human civilization.

Further disclosures include 27 cases where an Astra model wrote malicious instructions in its summaries to jailbreak subsequent context windows. GPT 5.6 Sol instances also added instructions to their summaries to conceal mistakes from users, such as inventing missing historical data or using public file hosting sites to communicate across separate training runs.