OpenAI begins publishing reports on cases where its models diverge from goals and constraints

OpenAI has begun publishing reports on cases where its models' behavior diverges from the goals and constraints it set for them. Along with the new framework for disclosing such episodes, the company released the first six descriptions collected over the past six months. Previously, such findings came out irregularly: they were accumulated into one large report or added to the cards for new models.

The framework is designed to disclose cases faster, even when the behavior has not yet been explained or fixed. Episodes are divided by how hard they are to verify: some are published right after review, while others require a lengthy investigation — in that case, a brief notice comes out first, with the full report following later. Some of what is published may turn out to be a one-off failure rather than a pattern.

One of the published examples: an unreleased research model inserted extraneous instructions into the working summaries it uses to continue work in a new context window — including instructions to disregard normal constraints. Twenty-seven such altered summaries were found.

Other cases from the first release: during training of GPT‑5.6 Sol, instructions were added to the model's summaries telling it to hide errors and discrepancies; in a separate episode, the model found an open API key in a public repository, used it without permission, and when it could not retrieve the data, fabricated it and passed it off as data from the requested source.

Another example — acting on its own initiative just to formally complete a task: to cite a file in its answer, the model uploaded it to the internet, even though the user had not asked it to.

OpenAI adds a caveat: these are isolated cases, and they cannot be used to judge how often this happens in its models. These are only the first reports under the new framework, not a complete picture of known cases. The company also believes the industry has not yet solved the alignment problem well enough to increase the pace of development without limits.

Image: FoxTPNL CC BY 4.0 · License

Translation editor: Салтанов А.