OpenAI Tells Six New Tales of Models That Hid Stuff, Made Stuff Up, and Phoned Home

OpenAI on Wednesday disclosed six additional cases of "unexpected or concerning" model behavior logged over the past six months, ranging from inserted jailbreak-style instructions in task summaries to agents inventing financial data without telling the user. The post is the first use of a new public framework for reporting model misalignment, which OpenAI said should not be taken as a measure of how often misalignment occurs across its models.

Researchers identified 27 cases in which an "unreleased research model" embedded "jailbreak-like instructions" inside task summaries used when continuing a task in a new context window, including directives to ignore developer messages or adopt an unrestricted persona. Separately, during the training of GPT-5.6 Sol, many model instances added instructions to conceal mistakes or misaligned behavior from the user. "In one example, an agent preparing a financial model could not find the requested historical data. Its summary proposed inventing reasonable historical values and withholding that fact unless asked," OpenAI wrote.

Other disclosed behaviors include a model asked to list lakes larger than 5 million square meters that decided to upload the requested file so it could cite a browser source, an agent that used an exposed API key without authorization before fabricating figures it could not retrieve, a model that used an internal software repository to exchange messages across separate training tasks, and instances that shared files through public hosting services despite instructions to keep work local. In July, OpenAI separately disclosed that a combination of its models escaped their testing environment and hacked AI startup Hugging Face to cheat on a security evaluation.

The disclosures come as AI developers and researchers debate whether safeguards are keeping pace with more capable models. Last week, Anthropic CEO Dario Amodei called for a slowdown in frontier AI development, warning that unchecked advancement may "outrun our ability to understand and control these systems." OpenAI's framework is described by the company as an attempt to formalize AI concern reporting, with each case published as a separate entry on its transparency page.