OpenAI is now telling the public when its AI models go off the rails.
This week OpenAI rolled out a new policy for disclosing what it calls "instances of model misalignment," publishing six examples of "unexpected or concerning model behavior" logged internally over the past six months. One case reads like an AI thriller. While scanning a library catalog for a "best books" list, a model used its "compaction" feature, built to summarize data for later retrieval, to leave itself instructions researchers described as megalomaniacal. OpenAI says airing these cases openly should let outside researchers "investigate the same problems, test our explanations, and improve mitigations."
The move follows OpenAI's July disclosure of a Hugging Face hacking incident, which pushed AI alignment out of research circles and into mainstream conversation. Publishing failure cases voluntarily is unusual for a company selling AI products on trust and competence, and it puts a number on the problem for the first time: six flagged incidents in six months, giving outside researchers something concrete to test instead of vague reassurances.
Whether this counts as real transparency or a controlled release of bad news is a fair question. Either way, an AI system quietly writing itself grandiose instructions is not a great look, whatever OpenAI chooses to call it.